Review row association independently from character accuracy

Correct text can be attached to the wrong record

Check which record each field belongs to so the spreadsheet supports real decisions rather than just looking like an accurate transcription.

In ParseMill, select an extracted invoice field to inspect its source, correct the value and export the reviewed workbook. Try your own invoices as well as the interactive sample. Compare upload, checking and correction time with your current way of building the same spreadsheet.

Define the unit of review as a field attached to a particular record. Comparing an unordered list of values can conceal a shift, especially when many results repeat or the same set of values appears on adjacent pages.

Work through a shifted-row fixture

Every source number appears somewhere in this invented output, but none is attached to its correct sample. An unordered value-set comparison would pass while a keyed field comparison would fail all three result assignments. This is why record identity must be part of the acceptance rule.

Microsoft: Document layout analysis represents table cells with row and column indices. These structural relationships are distinct from the recognized text and should be checked as such.

Test associations directly

1. Prepare expected records keyed by source document and sample identity. Retain source row locations for identifiers that are not unique.
2. Compare each observed field to the expected field on the same record. Do not match values by nearest text occurrence or by sorting both columns independently.
3. Inspect the source around every mismatch, including the preceding and following rows. Determine whether the error is character recognition, a row shift or a wrong key.
4. After repair, rerun the keyed comparison over the affected table, including page boundaries and repeated-value runs.

scikit-learn: Metrics and scoring emphasizes matching a metric to the target. A character-level score cannot stand in for a record-association score when the task depends on correct sample relationships.

Keep failures diagnostically separate

If A1’s source 11 is read as 17 on the correct row, that is a transcription failure. If 22 is read correctly but assigned to A1, that is an association failure. Both are wrong outputs, but separating them helps identify whether to improve recognition, layout handling or the combining step.

Check identifiers themselves as well. A corrupted sample key can make correct associated fields appear wrong or cause a join to the wrong expected record. Keep key validation visible in the review.

Use a demanding acceptance fixture

The shifted fixture should fail all three keyed result checks. Add a second fixture with repeated values and one distinctive transition to expose shifts that repeated zeros might hide. Record the actual outcomes of your own run rather than treating this illustration as product evidence.

When evaluating ParseMill, compare complete records with their source rows. Use both text fidelity and association checks before deciding whether an extracted table is ready for downstream analysis.