Wouldn't that just bias itself to the shape of the text extracted from the OCR against the shape of the raw text alone? It doesn't seem like it would be a great benchmark for estimating semantic accuracy?