A scanned contract can look perfectly clear on screen and still produce a translation with missing figures, broken table rows, or labels detached from their fields. The reason is simple: a scan is a picture of a document, not a document full of editable words.
Translating that picture takes more than sending a PDF to a translation tool. First, OCR has to recognize the text. Then the recognized text needs a careful check, a translation, and a layout pass that puts the translated content back into a useful form.
That sequence matters for any document where the page structure carries meaning: forms, invoices, contracts, certificates, manuals, and reports. A translation that gets the words right but attaches a figure to the wrong row can still be unusable.
What OCR does before translation¶
OCR, or optical character recognition, is software that reads text from an image and turns it into machine-readable characters. OCR is not translation: it identifies what a page appears to say, while a separate translation step changes that text into another language.
A scan may show familiar letters to a person, but a computer initially sees pixels. OCR analyzes those marks and tries to identify words, lines, and other page elements. The result can be searchable or editable, depending on the tool and the way the output is saved. The W3C technique for adding actual text to a scanned PDF describes OCR as a way to provide text that is otherwise only present as an image.
The distinction between recognition and translation is easy to miss. A file upload can accept a PDF and still fail to translate the words printed inside its page images. Google’s help page, linked in a discussion about translating a printed book, notes that text in images and scanned PDF pages may appear in the output without being translated. That kind of workflow needs OCR before translation.
A useful way to think about the process is to separate the document into three layers:
- The page image: the visible scan, including text, lines, stamps, signatures, tables, and marks.
- The recognized content: the text and structure OCR can extract from the image.
- The translated document: the target-language content placed into a readable page or file.
Those layers do not automatically stay in sync. OCR can misread a character. Translation can expand a phrase. Reconstruction can move a field or split a table. Each transition needs its own check.
The scan’s structure is also part of its meaning. A number under “invoice total” is not equivalent to the same number next to “tax.” A checked box and an empty box may have opposite implications. A handwritten amendment may qualify a printed clause. Layout is not decoration when the reader needs to know which label belongs to which value.
The W3C document calls its technique “Performing OCR on a scanned PDF document to provide actual text.”
The title captures the first task, not the whole workflow. Recognized text still needs translation and a layout pass before the translated file is ready to review.
Why scanned layouts break during translation¶
A scanned PDF does not contain editable paragraphs and tables in the same way a word-processing document does. The page stores an image, or a visual arrangement of content, so a translation system has to infer what belongs together before it can rebuild the page.
That reconstruction is where layout errors start. A system might recognize a heading, a paragraph, and a table as separate regions, but still put them in the wrong reading order. It might interpret a line through a form field as a character, or treat two columns as one continuous text block.
Even correct OCR does not solve the layout problem. A translation can be longer or shorter than its source. A label that fit on one line can wrap onto two. A translated table cell can push the next row down. A paragraph can spill onto a new page, leaving a signature line behind.
That is why “keep the layout” can mean different things:
- Keep the original page image: the source page stays visible as a background, while translation appears separately.
- Create a searchable PDF: the original appearance remains, with a hidden text layer that lets readers search or select recognized text.
- Rebuild an editable document: OCR-derived text and elements are placed into a Word file or another editable format.
- Create a translated page that resembles the source: translated content is laid out to follow the original structure, with adjustments where the new text does not fit.
Those outputs solve different problems. A searchable source PDF can be useful for finding text, but it does not necessarily give you a translated, editable page. A rebuilt Word document can be easier to review and revise, but it may not look exactly like the original scan. A close visual match may require layout work after translation.
The Library of Congress guidance on digitizing microfilm and OCR provides a useful reminder that recognition depends on the image being processed. Image quality, legibility, and the nature of the source affect what OCR can extract. A page that looks readable to a person can still contain details a recognition system cannot confidently distinguish.
A practical example makes the difference clear. Imagine a two-column supplier form with product descriptions on the left and quantities on the right. If OCR reads across the whole page instead of following each row, a quantity can end up next to the wrong description. Translating the resulting text accurately would not fix the underlying association.
The safer target is not “a translated PDF that looks close enough.” The target is a translated document in which a reviewer can connect every heading, value, field, and note to the right part of the original. Visual similarity helps, but semantic structure comes first.
What to check before you run OCR¶
Good OCR begins with a usable source image. Before you translate a scan, inspect whether the page is upright, whether the text is clear, and whether important details are visible at the size you will process.
Check the image, not just the PDF filename¶
A PDF can contain sharp scans, poor scans, photographs, or a mixture of all three. The file extension tells you little about whether the text is readable. Open each page and look at small print, punctuation, stamps, footnotes, and the edges of tables.
Look for the problems that make characters ambiguous:
- Blur: letters such as “e” and “c,” or digits such as “3” and “8,” may become hard to tell apart.
- Glare or shadows: a bright patch can erase part of a word, while a shadow can obscure a line.
- Rotation or skew: a tilted page can disrupt line detection and reading order.
- Small print: footnotes and dense tables can be harder to recognize than large headings.
- Cropping: a missing first or last character can change a name, code, or amount.
- Compression damage: rough edges and blocky artifacts can distort punctuation and thin strokes.
A few pages deserve extra attention even when the rest of a batch looks fine. A cover letter may be crisp while an inserted receipt is blurry. A contract may be typed, but the signature page may include handwritten dates or notes. Check each page type rather than judging the whole file by its cleanest page.
Separate print from handwriting¶
Handwriting is a different recognition problem from printed text. A typed date in a standard field may be straightforward to extract; a handwritten correction in the margin may not be. A clean-looking scan does not make handwriting reliably readable.
Mark uncertain handwritten areas for human review rather than letting an OCR system silently guess. That includes signatures, initials, handwritten amounts, corrections, checkmarks, and notes in margins. If a handwritten element changes the meaning of a clause or form, the reviewer needs the original image beside the extracted text.
Decide what must survive¶
Before processing, list which layout details matter to the person who will use the translation. For a form, the link between a label and a filled-in value may matter more than matching the original font. For a contract, clause numbering and cross-references may matter more than identical line breaks.
A quick preflight list can help:
- Keep the page order and section headings.
- Preserve table rows and the association between each value and its label.
- Keep checkbox states and form fields visible.
- Mark stamps, seals, signatures, and handwritten notes for review.
- Keep document identifiers, dates, and amounts easy to compare.
- Decide whether the final file must be editable, searchable, or visually close to the source.
Typical situation: a translated invoice needs the supplier name, invoice number, date, line items, tax, and total to remain easy to find. The reviewer does not need the same font as the scan, but they do need to see which amount belongs to which row.
A mistake at preflight can create avoidable rework. If the final user needs an editable file but the workflow only produces a flat page image, the translation may be readable but still hard to revise. If the user expects a near-identical visual copy, a plain text translation will not meet the brief.
How to review OCR text before translating¶
OCR proofreading compares recognized text with the scanned page before translation begins. The goal is to catch extraction errors while the source language is still visible, not to discover them after a translator has already carried them into the target language.
Start with a page-by-page comparison. Keep the scan open beside the OCR output and follow the reading order. Check headings, paragraphs, table rows, and fields as distinct units. A wall of extracted text makes it harder to notice when a value has detached from its label.
Then review the high-risk details first:
- Names and organizations: compare spelling, spacing, and any initials.
- Dates: check the day, month, year, and punctuation.
- Amounts and percentages: verify every digit, decimal mark, currency sign, and minus sign.
- Reference numbers: compare invoice numbers, contract IDs, account codes, and similar identifiers character by character.
- Checkboxes and marked fields: confirm whether each box is selected, empty, crossed out, or unclear.
- Tables: compare each entry with its row and column heading.
- Footnotes and marginal notes: check whether notes were picked up and connected to the right passage.
Those categories deserve attention because a small recognition error can change business meaning. A misread digit in a total is not a spelling mistake. A misplaced checkmark can reverse a response. A dropped minus sign can change the direction of a figure.
Do not limit the review to obvious words. OCR can confuse punctuation, omit a thin line, merge nearby columns, or insert text from a stamp into a paragraph. A recognizable sentence can still contain a wrong reference number or a missing qualifier.
For structured records, compare by region rather than reading the OCR output from top to bottom. Check the header, body, footer, tables, and handwritten additions separately. When a document contains repeated forms, compare the same fields across pages. A field that is empty on one page and filled on another may be meaningful, not a recognition gap.
A quality score can help decide which pages deserve additional review, but a score is not proof that the text is correct. A page can be easy to read overall and still contain one misrecognized character in a critical amount. Treat any automated quality signal as a way to prioritize, not as a replacement for comparing the actual text.
The Library of Congress OCR guidance is titled “Digitizing Microfilm and Optical Character Recognition (OCR).”
The pairing is a useful reminder: recognition starts with digitized image material. If the image does not show a detail clearly, OCR cannot reliably recover it just because the PDF opens without an error.
After correcting OCR errors, keep a record of uncertain passages. A reviewer can use a simple marker such as “unclear in source” rather than guessing. That note should remain visible through translation, especially when a number, name, or handwritten amendment could not be read with confidence.
Compare the main ways to translate a scanned document¶
The right method depends on what you need to deliver: a quick understanding of the content, an editable translated file, a close visual match, or an officially certified translation. No single route fits every scan.
| Route | What happens | Layout outcome | Main trade-off |
|---|---|---|---|
| Manual transcription and translation | A person reads the scan, types or transcribes the content, then translates it | Layout can be rebuilt deliberately | Requires careful manual work, especially for tables and repeated forms |
| OCR, then translation | OCR extracts text, a reviewer checks it, and a translation tool translates the corrected content | Structure may need to be rebuilt separately | Faster text processing does not guarantee a faithful page reconstruction |
| Document translation with OCR and layout handling | A document tool recognizes a scan, translates it, and generates an editable output | Can retain useful page structure, depending on the file and scan | Complex pages and poor scans still need review |
| Human translation and layout production | A translator works from the source and a layout specialist prepares the final file | Best suited to detailed visual control | More coordination and specialist work |
| Certified translation service | A qualified human translator handles the translation and certification requirements | Depends on the service and receiving authority’s rules | Certification is a separate human step, not an automatic feature of OCR |
A simple text extraction can be enough when the reader only needs to understand a letter or find a clause. A contract, statement, or form used for a decision often needs a more careful workflow, because the reader must be able to trace each translated value to its source.
When a translation bureau fits¶
A translation bureau can coordinate translation, document handling, and layout work. That can help when the scan is complicated, when someone needs to check the source against the output, or when the final file must meet a specific visual or certification requirement.
Before handing over a batch, explain what “keep the layout” means for your use. Ask whether the deliverable will be an editable document, a translated PDF, or a page designed to resemble the scan. Clarify whether tables, stamps, handwritten notes, and images will be recreated, carried over, or flagged for review.
A bureau is not automatically the right choice for every file. A clean scan with a straightforward layout may not need extensive desktop-publishing work. A faint, handwritten, or heavily degraded scan may need direct human interpretation rather than a routine OCR pipeline.
A useful brief includes the target language, intended reader, required file type, and any must-preserve elements. Mention whether the translation is for internal review or for submission to an authority. Requirements for certified translations depend on the recipient, so ask that recipient what form of certification it accepts rather than assuming a translated file is enough.
When a freelance translator fits¶
A freelance translator can be a good fit when you can supply clear files and agree on the expected output. Freelancers may also work directly with you on terminology, names, and document conventions, which is useful for recurring business documents.
The key question is who handles OCR correction and layout reconstruction. Some translators work from text supplied by the client; others take on document preparation as part of the job. If you send a scan, agree on who checks the recognized text against the original and who verifies the final page layout.
Ask how the translator will handle unclear content. A reliable workflow marks unreadable words, asks for clarification where needed, and avoids silently inventing a reading. For a table-heavy document, ask how row and column relationships will be checked after translation.
A freelancer may not have the tools or time to recreate a complex original page exactly. If the output must match a corporate template or preserve detailed design elements, layout work may need to be scoped separately.
When an online document tool fits¶
An online tool can handle a document-oriented workflow when the scan is legible and you want a translated file you can review and edit. The important difference from a basic text box is that the tool needs to process the document structure, not just return a block of translated words.
How it works in ChatsControl: you upload a scanned PDF or supported image, the scan is read and rebuilt as a formatted Word document, and the translated document follows that structure. Tables are rebuilt as tables, fields are placed beside their labels, and visible page elements such as stamps or signatures can be carried into the copy. The output still needs review, and the tool is not built for handwriting or very poor-quality scans.
A document tool is not a full translation project-management system. Teams that rely on translation memories, vendor workflows, or detailed segment management may need a separate CAT tool or TMS. A document translator solves a narrower problem: moving a file through recognition and translation while keeping it reviewable.
| Need | Online document tool | Human-led bureau or freelancer |
|---|---|---|
| First draft from a legible scan | Can automate recognition and translation steps | A person reads and translates the source |
| Layout | Can rebuild common structure into an editable output | Can handle detailed page design when scoped |
| Unclear handwriting | Needs human interpretation | A person can inspect and query the source |
| Terminology choices | May support a brief or terminology instructions | Can discuss choices directly with the translator |
| Certification | Requires a separate human service where needed | Can be arranged if the provider offers the required service |
| Final responsibility | Reader still needs to review important content | Translator and reviewer can check the output against the original |
A practical OCR-to-translation workflow¶
A reliable workflow separates source preparation, recognition, correction, translation, and layout review. Combining everything into one upload can save clicks, but it does not remove the need to check each stage.
Step 1: Define what the translated file must do¶
Write down the reader’s actual need before choosing a tool. Do they need to understand the content, edit the translation, submit it to an authority, or preserve a near-identical page design?
The answer changes what counts as success. An internal review copy can use clear headings and tables without matching every visual detail. A form intended for reuse may need editable fields. A document for an official submission may need a certified translation rather than an AI-generated draft.
Step 2: Prepare the scan¶
Check orientation, contrast, cropping, and legibility. Separate pages that have different conditions, such as a clean typed contract and a blurry receipt inserted at the end. Keep the original file untouched so the reviewer can compare it with any corrected or rebuilt version.
If an important detail is faint, don’t assume image processing will restore it. Ask for a better scan or a clearer original where possible. Mark any area that remains uncertain before translation begins.
Step 3: Run OCR and keep the page association¶
The OCR stage should produce readable text while preserving enough information to connect that text to its page and region. A plain text dump may be enough for a short letter, but a table-heavy document needs some record of row and column relationships.
Some OCR systems can detect more than character strings. Document-focused systems may recognize blocks, paragraphs, lines, words, symbols, and table regions. That structural information is useful when rebuilding a page, but it does not by itself prove that the recognized words are correct.
Step 4: Correct the OCR output¶
Compare the text with the page image. Fix errors in names, dates, amounts, identifiers, selected boxes, and table cells before the translator or machine-translation system sees the text.
Don’t “correct” unclear text by guessing. Mark uncertain passages and ask the document owner or source issuer for clarification when the meaning matters. When clarification is impossible, keep the uncertainty visible in the translated deliverable.
Step 5: Translate with context¶
Give the translator or tool the intended audience, subject area, and terminology preferences. A repeated product name, legal term, or internal label should not change from page to page without a reason.
Preserve identifiers exactly unless the document’s convention calls for a localized format. Dates may be formatted differently in another language, but the underlying date must stay the same. Review names, amounts, and codes separately from prose so formatting changes do not conceal an error.
Step 6: Rebuild the layout¶
Place translated text back into the document structure. Keep headings attached to their sections, table entries in their correct rows, and fields next to their labels. Make room for text expansion rather than shrinking type until the result becomes difficult to read.
A translation may take more space than the source. Adjust line breaks, cell widths, or page flow where needed, and check whether footers, page numbers, and signature blocks have moved. Layout reconstruction should make the content usable, not merely mimic the source at any cost.
Step 7: Review the final file against the scan¶
Compare the translated document with the source page by page. Check that no source text was omitted, no translated text was assigned to the wrong field, and no table row has shifted.
A bilingual review view can make this pass easier because the reviewer can compare source and translation together. In ChatsControl, the source and translated text appear side by side, with translated text highlighted and the original available on hover. That helps a translator inspect a draft, but it does not replace visual checking of the final rebuilt page.
Common failure points and how to avoid them¶
Most avoidable errors come from treating OCR, translation, and page design as a single operation. A process can produce fluent target-language text and still fail because it recognized a source number incorrectly or rebuilt a table in the wrong order.
A scan looks readable, so nobody checks OCR¶
Human readers use context to fill in gaps. OCR software can make a confident-looking guess where the image is ambiguous. A reviewer should compare the extracted text with the scan, especially for figures and identifiers.
A second check is particularly useful for values that affect payment, eligibility, or contractual obligations. Ask someone who can read the source language to verify those entries against the page image. Don’t rely on a polished translation as evidence that the source was recognized correctly.
A table becomes a list of unrelated values¶
Tables encode relationships through rows, columns, headings, and spacing. OCR may recover the words but lose the structure that explains which value belongs to which heading.
Review tables by row and column, not only by reading the extracted text in sequence. Check whether the translation keeps the header associated with the right column and each entry with the right row. If a table cannot be reconstructed reliably, use a human layout pass or mark the content for a separate review.
The translation changes line length and pushes content around¶
Translation changes the amount of space a passage needs. A narrow cell may become too small, an address may wrap, or a signature line may move to another page.
The layout reviewer should check the translated document after all text has been inserted, not just approve a blank template. Confirm that page breaks, headings, table rows, footnotes, and signature blocks remain understandable. A visually similar page is not successful if a shifted label now points to the wrong field.
Handwriting gets treated as certain text¶
A system may be able to read some handwriting, but a plausible transcription is not proof of accuracy. Initials, dates, annotations, and handwritten corrections can be critical even when they occupy only a small part of the page.
Route uncertain handwritten material to a human who can compare the image with the surrounding document. If a mark cannot be deciphered, preserve that uncertainty rather than converting it into a confident but unsupported word.
Official-use expectations are discovered too late¶
An AI-translated document is not automatically certified because it looks formal or carries the original stamp image. A receiving authority may require a particular translator, certification, or document format.
Check the recipient’s requirements before commissioning the translation. A human certified or sworn translator may need to review and certify the result. ChatsControl’s certified-translation route involves a human translator and is not an instant automated step; the service details are available through certified translation.
Confidentiality is not part of the workflow brief¶
A scan can contain personal information, commercial terms, or account details. Before uploading it to an online service, check whether its handling fits your organization’s requirements.
Ask who can access the file, how long it remains available, and whether the tool offers retention controls. ChatsControl uploads are set to auto-delete after 30 days by default. Organizations with strict on-premise or offline requirements should not assume a cloud service meets those conditions.
FAQ¶
How do I translate a scanned document and keep its layout?¶
Run OCR to recognize the text, compare that text with the scan, translate the corrected content, then rebuild the document and check the result page by page. OCR extracts text; a separate layout step places the translated text back into a usable structure.
Can Google Translate translate text in scanned PDF pages?¶
Google’s documented file-translation feature says text in images and scanned PDF pages may appear in the output but will not be translated. For that workflow, use OCR first and translate the resulting editable text or document.
What scan quality is needed for OCR translation?¶
The text should be legible, correctly oriented, and detailed enough to distinguish letters, punctuation, and marks. Blur, glare, small type, rotation, handwriting, and degraded images can make recognition less reliable.
How do I check OCR errors before translating a business document?¶
Compare the recognized text directly with the scan. Check names, dates, amounts, reference numbers, checkboxes, and table cells carefully, and correct recognition mistakes before translation.
Why does a translated scanned PDF look different from the original?¶
OCR recognizes text, but it does not automatically place the translation into the source design. Translated text can take more or less space, so the file needs reconstruction and a final visual review.
Can I use an OCR translation as a certified translation?¶
OCR or AI translation is not automatically certified. Ask the receiving organization what it accepts, then arrange the required human translation and certification if needed.
Is OCR accurate enough for contracts and business forms?¶
OCR can help make a scan searchable and editable, but its output should not be treated as error-free. Verify high-impact information against the original before relying on the translated document.