Improving Image-Sentence Embeddings Using Large Weakly Annotated Photo Collections