Visually Grounded VQA by Lattice-based Retrieval