Towards Robust Visual Information Extraction in Real World: New Dataset and Novel Solution