Converting PDFs to Markdown is now routine preprocessing for search and language models, yet scoring the numbers in the output is easy to get wrong. We evaluate MinerU, a plain-text OCR baseline (NDLOCR-Lite), and a deterministic post-processing pipeline (anchor insertion, footnote reordering, fixes of look-alike characters in fixed contexts, gated formula repair using two independent OCR partners, and heading reconstruction) on 48 public Japanese pages (three documents), text-layer and rasterized. Our pre-specified number-and-unit F1 favored the OCR baseline (0.575 vs 0.397), and a disagreement check between the two engines seemed to flag about half of the converter's number errors (54.6%). Both were artifacts. The scorer counted digits in the converter's notation (image-link paths, HTML attributes), which plain text lacks. With the notation removed (a post-hoc check), the F1 scores tie (0.575 vs 0.594), and the check flags only 13% of number errors, because most errors are misses that both engines share. What remains: the OCR baseline has a lower body character error rate (0.012 vs 0.053 rasterized). The pipeline raises extra-number recall at page level (+0.145; the document-level interval touches 0), mostly by restoring page numbers, but lowers heading-level agreement (0.133 to 0.061) and leaves spurious formulas unchanged; its formula repair never fires. A pre-specified precision gate for its flags (0.80, human-judged) is met at 0.929, yet the post-processing hypothesis fails on its improvement items. We recommend scoring numbers after removing notation and checking critical figures against the source. By the pre-specified criteria, the study stays a technical report.
翻译:暂无翻译