On Web-based Visual Corpus Construction for Visual Document Understanding