When processing a document type that comes in a high variety of layouts, what is the recommended data extraction methodology?