Improving Pre-trained Vision-and-Language Embeddings for Phrase Grounding