What's so hard about PDF text extraction?