Improving Information Extraction from Images with Learned Semantic Models