Key takeaway
Searchable text and verified field values are different deliverables. Review the errors that can change a decision before scaling extraction.
Choose the question the extracted text must answer
Making a report searchable can be useful even when some words are wrong. Turning the same report into an inspection decision requires a stricter review of measurements, limits, units, dates and negation. Write the intended use before selecting an accuracy measure. A high average character score can conceal the single altered digit that reverses a pass/fail conclusion.
The National Archives describes OCR as machine-generated transcription that is not always accurate and provides a process for edits and validation. That Catalog practice supports checking extracted text against the original; it does not establish a commercial tolerance or certify a private archive. Choose your own critical-field review with the people who understand the document.
A completed critical-field review
This hypothetical inspection archive contains scanned reports. The intended evaluation needs recorded measurements and the inspector’s disposition. An internal reviewer compares selected fields with the source page, preserving the image identifier and region alongside the extracted value. Each row below is invented to demonstrate the decision; it is not a measured error rate for any OCR product.
The review distinguishes a readable correction from an unresolved reading. If the original is blurred, two reviewers agreeing on a guess does not create observed evidence. Keep the field unresolved, retain the reason and decide whether that report can answer the proposed question.
| Source-page reading | OCR output | Consequence | Decision |
|---|---|---|---|
| 0.08 mm | 0.8 mm | Tenfold measurement change | Correct after source verification; retain old extraction |
| Not accepted | Accepted | Disposition reversed | Require explicit negation review |
| Lot B10 | Lot BIO | Wrong record link | Verify identifier against source context |
| Unreadable handwritten digit | 7 | False certainty | Keep unknown; exclude from numeric test |
Separate page coverage from field correctness
First reconcile expected pages with captured pages: a perfect transcription of page one cannot recover a missing signed continuation. Check orientation, cropping, front/back capture and whether attachments were included in the approved source scope. Then assess the critical fields on the pages that actually exist. These are separate denominators and should have separate results.
For an illustrative internal check, suppose 100 expected pages produce 98 captured pages, and 200 reviewed critical fields contain six confirmed extraction errors and four unresolved readings. Report the two missing pages, six errors and four unresolved readings separately. Do not call all ten fields errors when their source truth is not known, or quietly remove the unresolved fields from the reported review.
Choose a review rule by field consequence. An internal search experiment may tolerate misspelled descriptive words while requiring manual verification of every measurement used as a reference answer. State that selected rule and its coverage. Do not extrapolate a targeted numeric-field check into an accuracy claim about all pages, languages or handwriting in the archive.
Keep correction provenance beside the value
Retain the source image identity, page or region, extraction tool/version, raw output, corrected output, reviewer, date and reason. W3C’s provenance overview identifies entities, activities and people as the context for assessing how data was produced. A compact correction ledger applies that idea without claiming that documentation alone proves accuracy.
Freeze a release before external evaluation. If the extraction settings change, rerun the affected check and identify the new package version. Keep a corrected transcription distinct from a human interpretation of the document. Replacing “0.8” with the readable “0.08” is a transcription correction; deciding that a measurement was acceptable under a particular specification is a separate assertion.
Stop where the source cannot support the use
For this example, searchable discovery can proceed internally while the numeric evaluation holds reports with missing pages or unreadable critical digits. The next work is targeted recapture or review, not a full-archive cleanup justified by a vague average accuracy claim. Set a bounded review budget and an owner for each unresolved document class.
Before any actual sample, review source rights, confidential annotations and personal information in both the image and extracted text. A clean text export does not clear the underlying page. Use readiness to track the remaining evidence and inventory to describe the approved document category. Share a metadata description first; the OCR ledger belongs in the later separately approved evaluation.