In-depth practical guides

Check OCR extraction before citing a document

Move from a scanned page to verified information: check characters, tables, reading order and languages before searching, calculating or citing automatically recognized text.

Updated :

Archive consultation with a microfilm reader and historical documents.

AI-generated illustration

Does the extracted text preserve the document's meaning?

Selectable text may contain initially invisible mistakes: a lost negation, substituted digit, mixed column or split name. OCR helps discovery, but the source page remains the comparison point. Distinguish extraction useful for finding a passage from extraction reliable enough for a quotation or calculation. No universal accuracy rate replaces an assessment tied to the intended use.

  1. Keep the source

    Record document, provenance, retrieval date and pagination. Retain the page image and extracted text separately with version identifiers. A new extraction should not overwrite previously used evidence. Check that access and reuse conditions fit the intended work.

  2. Select passages to inspect

    Include a simple page, table, note, narrow column and the scripts actually present. Add every value supporting the conclusion: amounts, dates, units, names and negations. Easy pages alone poorly represent a varied corpus; state the chosen coverage.

  3. Compare decisive characters

    Read the image and extraction side by side. Check zero versus O, one versus I, decimal separators, minus signs and references. Review numbers with their units and periods. Mark corrections in the transcription without silently altering the source file.

  4. Reconstruct structure

    Follow headings, paragraphs, columns, notes and cells. A table row is more than a word sequence: connect each value to its header. Check tables continued across pages. Retain the original page or region for each reused datum.

  5. Test search and languages

    Search for a name, reference and phrase in each script. Inspect accents, Arabic characters, punctuation and mixed-script ordering. Separate recognition of characters from translation of meaning. Avoid automatically correcting a proper name with a general dictionary.

  6. Decide the permitted use

    Classify passages as checked, corrected with a record, or unreliable. Partial extraction may support discovery while remaining unsuitable for calculation. Reread quoted passages directly and provide a stable locator. After changing the engine or settings, repeat difficult passages rather than merely comparing overall volume.

Reading workshop

A good score does not validate the identifier

Case DEMO-O12 separates character agreement and block order before citation or joining.

Fictional data and records for learning: none of these results comes from a live service query.

Column order and the identity of a code

The page should be read left column then right column. OCR alternates the blocks and replaces two zeros in the code with O letters; character agreement does not cover that ordering defect.

If the table exceeds the screen, scroll horizontally; arrow keys work when the region is focused.

Column order and the identity of a code
StepScenario dataDiagnosisDecision
Two-column imageLeft A1,A2; right B1,B2Expected order A1,A2,B1,B2.Retain image and region references.
Raw extractionA1,B1,A2,B2Columns alternate and mix the argument.Restore order by reading the image.
Decisive identifierImage AB-0010; OCR AB-OO10Two zeros become O letters.Correct with a record before reuse.
Aligned character sample2 errors in 10,000; agreement 99.98%This fictional score does not measure block order.Read every identifier supporting the conclusion.

Can AB-OO10 be cited because agreement is 99.98%?

Open each proposition to compare its diagnosis. More checks may still be needed.

Yes, the score is nearly perfect.

No: two errors can change the key being used.

Yes, after restoring column order.

No: structure and characters require separate checks.

No; read the code from the image and record the correction.

Appropriate. Also check the whole passage; keep raw text separate from corrected transcription.

Filled fictional record

File: DEMO-O12
Image: AB-0010; raw OCR: AB-OO10
Raw order: A1,B1,A2,B2
Reviewed order: A1,A2,B1,B2
Sample: 10,000 characters; errors: 2
Correction: code and order with page references
Use: checked passage only

Attach the reviewed page and image region to the cited passage. Retain raw extraction, corrected transcription and the rationale for each code correction separately.

Documentation of the rules, separate from fictional examples

Compare other situations and exercises

A negation disappears

The extraction allows something the printed page excludes. Compare the whole sentence, its note and context. Record the correction and return to the source when citing it; general word similarity does not protect meaning.

Two columns merge

Headings are searchable, but paragraphs alternate between columns. Reconstruct order from the image before summarizing. Then test a full-sentence query: words present in the index need not belong to the same passage.

A total looks inconsistent

A value lost its sign or shifted column. Check each useful cell, its headers and unit. Recalculate only with verified values and document unreadable cells instead of filling them by intuition.

Guided reading of an excerpt

A lost separator multiplies an amount by one hundred

In this fictional exercise, A transcribes the value read in the source image and B is the OCR result. Page and table identify the area to revisit.

Fictional exercise: identifiers, values and dates illustrate the check. This excerpt is not a live collected response.

  1. Find the cell

    Connect page 7, table 2, heading and EUR unit. Do not check a number without context.

  2. Compare characters

    A uses a decimal comma and a grouping space. B lost both; 125000 does not mean 1 250,00.

  3. Correct the transcription

    Keep raw output and a separate correction linked to the cell. Recalculate results that used the wrong scale.

Fictional comparison: A = image reading; B = OCR

p. 7 / table 2
A   1 250,00 EUR
B   125000 EUR

What to inspect

Accepted value: 1 250,00 EUR
The raw 125000 EUR would be one hundred times larger.
Scope of the check
One verified cell validates neither the whole table nor every page. Also check decisive negations, names and units.
Is adding a decimal separator to every OCR number a reliable correction?

Reasoning

No. Conventions and errors can differ by column or page. Justify every consequential correction from the source area; a blind global rule can introduce new errors.

Official references : Tesseract — Improving the quality of the output

READING / ANNOTATED CASE

Two character errors, two critical fields to review

Compare two fictional strings while preserving case, signs and all three spaces. The reference has 18 ASCII characters and four tokens. For this example, tokens are split on spaces with punctuation retained: this is the declared WER convention below, not a universal setting.

Fictional example · no observation of a service or client.

On small screens, scroll the table horizontally.

Two character errors, two critical fields to review
CheckExample valueInterpretation
Reviewed referenceREF A01 -12.50 EUR18 characters including spaces; four tokens.
Raw OCR outputREF A0I 12.50 EUR17 characters; digit 1 became I and the minus sign disappeared.
CER: character errors2 / 18 × 100 ≈ 11.1 %One substitution and one deletion, divided by reference characters.
WER: this example’s convention2 / 4 × 100 = 50 %Two substituted tokens out of four: the identifier and signed amount.
Fitness for useA01 ≠ A0I ; -12.50 ≠ 12.50The identifier can point to a different object and the sign reverses the value. Return to source regions before matching or calculation.
Is it possible to claim 88.9% reliability?

This complement of CER measures neither document truth nor the probability that each field is correct. Both useful fields are altered here despite the strings’ similarity.

Why document conventions?

Stripping punctuation before scoring could erase the sign error. Compare extractions using the same rules, then separately examine critical fields and reading order.

References accessed on

Worked exercise

A lost sign reverses the conclusion

Fictional example: the values and documents teach the method and do not describe a real organisation.

The column represents percentage change. The automatic summary calls it an increase.

If the table exceeds the screen, scroll horizontally; arrow keys work when the region is focused.

Starting information
SourceRead text
Table image−12,5
OCR extraction12.5

Your checks

  1. Read the cell on the original image.
  2. Check sign, decimal separator and heading.
  3. Recalculate the conclusion from the corrected cell.
Read the answer

Answer and decision

The value is −12.5%, a decrease in this table. Changing a decimal comma to a point can be formatting; losing the sign changes meaning. Correct the summary and derived calculations too, while keeping the original extraction identifiable.

If the case changes

If a dash might be a missing-value code, inspect the legend and alignment; do not automatically treat it as a minus sign.

Record to retain

Keep page, cell location, readable image, raw OCR text and justified correction.

DOCUMENTED REFERENCE

ALTO: two confidence scales running in opposite directions

ALTO 4.4 defines word confidence WC and character confidence CC with opposite endpoints. These are OCR metadata, not measured accuracy for your corpus.

ALTO: two confidence scales running in opposite directions
ReferenceValue or formWhat it identifies
WC0 → 10 means unsure and 1 sure according to the engine; a high score does not replace reading the image.
CC0 → 9One number per character: 0 means sure and 9 unsure. Do not treat 9 as greater confidence.
HPOS / VPOSMeasurementUnitHorizontal and vertical positions in the declared measurement unit; do not assume pixels.
CONTENTStringRecognised content of a text unit; retain its page and region link.
ReadingOrderALTO 4.3+Explicit reading order introduced in version 4.3; check the version before expecting this element.

Implication for the record

Keep version, unit, text and coordinates together. Institutional profiles can differ from the general schema; BnF documents its ALTO variants. A missing score is not zero confidence.

Does WC = 0.99 establish 99% correct characters?

No. WC is engine confidence for a text unit, not an observed character-accuracy measure. Evaluate passages against a reference transcription and describe the sample rather than treating a score as certification.

Primary documentation

Documentation consulted on .

DOCUMENTED REFERENCE

A 1930 ownership mark does not date the book’s printing

This reproduction shows an opening of the Library of Congress Gutenberg Bible, with provenance marks on the inside cover. It comes from WDL7782 as relayed by Wikimedia Commons. Describe the work and this copy’s history separately.

Opening of the Library of Congress Gutenberg Bible: inside cover with bookplates, a Vollbehr collection label dated 1930 and a flyleaf.

Opening of the Library of Congress Gutenberg Bible: inside cover with bookplates, a Vollbehr collection label dated 1930 and a flyleaf.

Authentic documentary reproduction · Library of Congress · World Digital Library · Wikimedia Commons · Source · Reuse

Reproduction marked as public domain on the file record. Resized and converted to WebP without cropping. Photograph date unspecified.

A 1930 ownership mark does not date the book’s printing
ReferenceValue or formWhat it identifies
Biblia latinaMainz · c. 1455The Library of Congress exhibition places printing in Mainz, probably completed late in 1455.
LanguageLatinThe Bible’s text is Latin, the Vulgate; a collection label’s language is not the language of the entire book.
MaterialVellumThis copy is printed on vellum according to its institutional description.
Vollbehr1930Year the Library acquired the Vollbehr collection, also visible on the reproduction’s collection label.
WDL7782Identifier of the World Digital Library record cited in file provenance, not pagination.

Implication for the record

In OCR, separate work text, annotations, bookplates and added labels. Attribute each statement to its region and function. Digitisation or photography date is not supplied here and must not be inferred from 1455 or 1930.

What does “The Vollbehr Collection 1930” establish?

It is a collection mark, not a printing year. The institutional description independently documents the 1930 acquisition. A verifiable citation preserves that distinction and the precise copy: one volume’s ownership mark does not describe every copy of the edition.

Primary documentation

Documentation consulted on .

DOCUMENTED REFERENCE

PDF/A: a preservation profile is not an authenticity certificate

ISO 19005 parts specify preservation requirements for rendering and resources. This table compares documented profiles, not checks performed on visitors’ PDFs. Identify the exact profile before interpreting a conformance report.

PDF/A: a preservation profile is not an authenticity certificate
ReferenceValue or formWhat it identifies
PDF/A-1ISO 19005-1:2005The first part, based on PDF 1.4. Requirements support reproducible rendering; a .pdf extension does not establish this profile.
PDF/A-2ISO 19005-2:2011Based on PDF 1.7, with additional features and PDF/A embedding under its rules. This does not permit arbitrary attachment formats.
PDF/A-3ISO 19005-3:2012Allows embedded files in arbitrary formats. Container conformance does not make attachments themselves preservable, accessible or PDF/A-conforming.
PDF/A-4fISO 19005-4:2020A part 4 profile based on PDF 2.0, allowing embedded files in other formats. The profile name does not reveal their nature or quality.
pdfaid:part / pdfaid:conformance1 / BAn XMP declaration example for PDF/A-1b: part is 1 and conformance is B. Metadata states claimed conformance, not successful validation.
PDF/UAISO 14289The PDF accessibility standards family. A file can target both PDF/A and PDF/UA; PDF/A alone does not establish accessibility.

Implication for the record

Retain exact file, tested profile, tool and version, applied rules and results. OCR, reading order, attachments and signatures need separate checks. Format conformity certifies neither author, publication date nor claims.

Is a PDF/A with selectable text necessarily accessible?

No. Selection does not establish semantic structure, correct reading order or suitable image descriptions. Check relevant accessibility requirements and read with assistive tools.

Primary documentation

Documentation consulted on .

Record to retain

  • Source, version, page image and retrieval date.
  • Checked passages, scripts, layout and selection criteria.
  • Errors, explicit corrections and locations of reused values.
  • Permitted uses: search, citation or calculation, with limits and planned rechecking.

Official references

Related protocols

Checks before reaching a conclusion

  • Have decisive negations, signs, decimal separators and proper names been compared directly with the visible page?
  • Do table cells retain the correct heading and order after being copied into the working tool?
  • Does the citation point to the source page and identify corrections rather than only the extracted text?

Frequently asked questions

Is a high confidence score sufficient?

It can help prioritize review but cannot replace comparison of decisive passages. One error may reverse a sentence's meaning.

Must the whole document be transcribed?

Match the work to the decision. For a specific quotation, check the passage and context; for a calculation, verify every used cell.

Can a scan be improved after extraction?

Compare new extractions on the same passages. Keep the initial image and record transformations to distinguish the source from prepared files.

Does OCR translate a document?

No. Text recognition and translation are separate stages with different errors and checks. Keep the passage in its source language.

What steps should be controlled in priority?

The numbers, names, dates, and reserves of which the conclusion depends. A significant error may justify the examination of the tables or passages produced by the same recognition process.

DOCUMENTED REFERENCE

OCR: measure errors against an explicit reference and profile

Levenshtein distance counts the minimum operations needed to turn a reference into candidate text. It can examine a transcription or excerpt, but does not assess translation faithfulness or answer truth.

OCR: measure errors against an explicit reference and profile
ReferenceValue or formWhat it identifies
S / D / Ireference → candidateS replaces one reference unit; D removes one; I adds a unit to the candidate. Each costs 1. Swapping letters is not a special operation here.
N / MH+S+D / H+S+IN is reference length after declared processing. With H identical units, N = H + S + D; candidate length is H + S + I.
CER / WER (%)100 × (S+D+I) / NThe percentage rate is 100 × (S + D + I) / N and may exceed 100%: a versus aaa requires two insertions for one reference unit.
units / profilecode points · White_Space 17.0.0 · raw / NFCCharacter mode counts code points, not graphemes or UTF-16 units. Word mode splits on the 25 Unicode 17.0.0 White_Space characters without linguistic segmentation. Case and punctuation remain; NFC is optional.
tie-breakH → S → D → ISeveral alignments can share the minimum cost. Backtracking prefers H, then S, D and I. For ab versus ba, two substitutions are selected; deletion and insertion can have the same cost.
N = 0undefined rateFor N = 0, this explorer shows operations and leaves the rate undefined, even for two empty texts. JiWER 4 uses other empty-reference conventions: compare outputs only under the same contract.

Compare texts unit by unit

Original teaching example; no measurement of a real OCR engine.

At most 200 code points per text before and after NFC. Case and punctuation remain. Text fields use LF line endings.

Tools calculate in your browser without transmitting entered text.

Comparison calculated · WER · raw · N = 3 · E = 1

Reference units N
3
Candidate units M
4
Distance E = S + D + I
1
Error rate (%)
33.333333

Substitutions S : 0Deletions D : 0Insertions I : 1Identical units H : 3

Chosen alignment: minimum cost, then H, S, D, I priority during backtracking. Other optimal alignments may divide errors differently.

Alignment operation countsCounts S, D, I and H. The axis runs from 0 to max(1,N,M). These are unit counts, not percentages.040S0D1I3H
Counts S, D, I and H. The axis runs from 0 to max(1,N,M). These are unit counts, not percentages.
View alignment and operations

∅ means no unit. Units use JSON notation to expose spaces and line breaks.

View alignment and operations
StepReference unitCandidate unitOperation
1"le""le"Identical units H
2"chat""chat"Identical units H
3∅"noir"Insertions I
4"dort""dort"Identical units H

Texts actually compared, in JSON notation

{"reference":"le chat dort","candidate":"le chat noir dort"}

Formula and profile

N = H + S + D; M = H + S + I
E = S + D + I
N > 0 → rate = 100 × E/N
N = 0 → undefined
Original examples: units, profiles and edge cases

Example texts use JSON notation with Unicode escapes; those escapes do not define the counting unit.

Original examples: units, profiles and edge cases
Checked reference textCandidate textFormula and profileNEError rate (%)
"le chat dort""le chat noir dort"WER · raw3133.333333
"le chat dort""le chat noir dort"CER · raw12541.666667
"a""aaa"CER · raw12200
"cat"""CER · raw33100
"""x"CER · raw01Undefined
""""CER · raw00Undefined
"ab""ba"CER · raw22100
"\u00e9""e\u0301"CER · raw12200
"\u00e9""e\u0301"CER · NFC100
"Le chat.""le chat"WER · raw22100
"les chats""les chats"WER · raw200
"les chats""les chats"CER · raw10110
"un\u00a0chat""un chat"WER · raw200
"\u6587\u672c""\u6587\u4f53"WER · raw11100
"\u6587\u672c""\u6587\u4f53"CER · raw2150
"\u0634\u064e""\u0634"CER · raw2150
"\ud83e\udded""\ud83e\udded\ud83e\udded"CER · raw11100
"un\u200bchat""un chat"WER · raw12200

White_Space is fixed to Unicode 17.0.0; NFC uses the browser Unicode engine. Accents, case and punctuation are not removed.

Unicode data and license
White_Space data from PropList-17.0.0.txt. Copyright © 2025 Unicode, Inc.

UNICODE LICENSE V3

COPYRIGHT AND PERMISSION NOTICE

Copyright © 1991-2026 Unicode, Inc.

NOTICE TO USER: Carefully read the following legal agreement. BY
DOWNLOADING, INSTALLING, COPYING OR OTHERWISE USING DATA FILES, AND/OR
SOFTWARE, YOU UNEQUIVOCALLY ACCEPT, AND AGREE TO BE BOUND BY, ALL OF THE
TERMS AND CONDITIONS OF THIS AGREEMENT. IF YOU DO NOT AGREE, DO NOT
DOWNLOAD, INSTALL, COPY, DISTRIBUTE OR USE THE DATA FILES OR SOFTWARE.

Permission is hereby granted, free of charge, to any person obtaining a
copy of data files and any associated documentation (the "Data Files") or
software and any associated documentation (the "Software") to deal in the
Data Files or Software without restriction, including without limitation
the rights to use, copy, modify, merge, publish, distribute, and/or sell
copies of the Data Files or Software, and to permit persons to whom the
Data Files or Software are furnished to do so, provided that either (a)
this copyright and permission notice appear with all copies of the Data
Files or Software, or (b) this copyright and permission notice appear in
associated Documentation.

THE DATA FILES AND SOFTWARE ARE PROVIDED "AS IS", WITHOUT WARRANTY OF ANY
KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT OF
THIRD PARTY RIGHTS.

IN NO EVENT SHALL THE COPYRIGHT HOLDER OR HOLDERS INCLUDED IN THIS NOTICE
BE LIABLE FOR ANY CLAIM, OR ANY SPECIAL INDIRECT OR CONSEQUENTIAL DAMAGES,
OR ANY DAMAGES WHATSOEVER RESULTING FROM LOSS OF USE, DATA OR PROFITS,
WHETHER IN AN ACTION OF CONTRACT, NEGLIGENCE OR OTHER TORTIOUS ACTION,
ARISING OUT OF OR IN CONNECTION WITH THE USE OR PERFORMANCE OF THE DATA
FILES OR SOFTWARE.

Except as contained in this notice, the name of a copyright holder shall
not be used in advertising or otherwise to promote the sale, use or other
dealings in these Data Files or Software without prior written
authorization of the copyright holder.

Implication for the record

Original examples, not measured software performance. A low rate guarantees neither reading order nor correct names or amounts. Check the reference itself against the document. Browser text fields convert line endings to LF.

Why are OCR confidence and error rate different?

Engine confidence is its own estimate. CER or WER compares output against a chosen reference under a calculation profile. Preserve both with their methods; neither replaces checking an important passage.

Primary documentation

Documentation consulted on .

DEEPER READING / PRACTICE

Verify the passages supporting the conclusion

A digitized document has been converted into text. The control should focus on the names, dates, numbers and reserves that influence reading and then compare to the visual document.

Updated on .

Priority checks

Examine the quoted passages, tables and page ruptures. Release characters confusion, fused lines and notes whose order has been changed.

Interpretation pitfalls

A fluid text may contain a major digital error. A full text search may lack a name or a word that is misrecognized.

Working record to retain

The OCR journal combines passage, recognized text, verified reading and correction. The extract used in the conclusion is linked to their location in the original document.