Verify OCR text before citing a scanned research paper

A scanned paper can look clear in a PDF viewer while its copied text contains a different number. A missing zero, a lost minus sign, or a table value attached to the wrong heading can change what you record in a literature review. Finding the passage is only the beginning of checking it.

Optical character recognition (OCR) converts page images into machine-readable text. The University of Michigan Library explains that this automated transcription can be inaccurate when the scan, print, or font is difficult to read. Its plain text is therefore not a fully faithful copy of the original. Use OCR to locate evidence, then verify the material you intend to quote, summarize, or extract.

1. Keep the original and the transcription

You need the original PDF or page images, the extracted text, and somewhere to record corrections. A PDF viewer and a spreadsheet are enough. Save the source file unchanged. Give it a stable local identifier and retain its DOI or bibliographic record when available. If you rerun extraction, keep the earlier output with the tool name, date, and settings rather than silently replacing it.

Test one relevant paragraph by selecting and copying it into a plain-text editor. A page with selectable words can still have an inaccurate hidden text layer. Conversely, a page without selectable text may simply require OCR. Neither observation establishes whether the scientific claim is correct.

Record both the PDF page position and the printed page label. For example, the seventh page in a file might be labeled 114 because it belongs to a journal issue. Add a section, table number, or short passage locator so another reader can find the same material without repeating your search.

2. Inspect the passage that supports the claim

Open the page image beside the extracted text. Read the whole sentence, its surrounding paragraph, and any referenced table caption or footnote. Check the characters that determine the meaning, not just whether the transcription sounds fluent.

• Numbers and signs: compare every digit, decimal point, minus sign, inequality, and exponent used in the result.
• Units: retain the printed unit and its prefix. Check superscripts and subscripts before converting a value.
• Reading order: confirm that the sentence continues in the correct column and does not incorporate a running header or footnote.
• Tables: trace the value to its row label, column heading, and footnote. A correct number assigned to the wrong condition is still an incorrect extraction.
• Qualifiers: verify words such as “not,” “only,” and “approximately,” together with the population or operating conditions they qualify.

Tesseract's documentation identifies skew, image noise, and page segmentation as factors in recognition quality, and notes limitations with tables without additional layout analysis. Those are reasons to inspect difficult regions individually. They are not a universal error rate for every OCR system or every research paper.

3. Work through a decimal-point check

This is a deliberately constructed teaching example, not a measurement from a published paper or an observed output from an OCR engine. Suppose a methods page states that a fixture has a clearance of 0.05 mm. The deliberately corrupted transcript below says 0.5 mm. One missing zero makes the transcribed clearance ten times the source value: 0.5 divided by 0.05 equals 10.

The check has three parts. First, inspect the zero in the page image. Second, confirm that “mm” belongs to the clearance rather than another column. Third, reread the sentence to establish whether the clearance is specified, measured, or merely used in a model. Correcting the number does not answer that last question by itself.

For this fixture, record 0.5 mm as the raw transcription and 0.05 mm as the verified transcription. Give the reason “zero omitted after decimal point.” Keep your interpretation in a separate field, such as “specified fixture clearance.” Do not silently rewrite a quotation or claim to have validated the experiment.

4. Make the correction reviewable

A correction log needs enough context to reconstruct your decision. Use these fields in a spreadsheet or your evidence notes:

If a passage cannot be read reliably, mark it unresolved. Seek a better scan, an accessible publisher version, or clarification from the source owner. Do not use surrounding prose, a language model, or an expected result to fill in an unreadable digit. A plausible repair is still an inference.

After a correction, revisit any note, calculation, or draft sentence that used the earlier text. Link the verified passage to your evidence matrix, retaining the source location. If the value changed, recompute derived quantities and check their units rather than updating only the displayed quotation.

5. Decide what the check does and does not establish

The expected output is a verified passage with a traceable correction record. It is not a certificate that the whole PDF is accurate. Checking one paragraph does not validate untouched pages, equations, or tables. For a review, inspect every passage and value that materially supports an extracted finding or a draft claim.

An OCR correction also cannot fix an error in the original publication. If the page itself appears inconsistent, preserve its wording and record the discrepancy separately. Check for a correction notice or another relevant source before deciding how to use the result.

Finish with a claim audit: does the verified passage support the exact sentence you wrote, under the same conditions and with the same limitations? Faithful transcription is necessary for that comparison. It does not replace it.