How People Translate Quickly | Source Cleanup: Fix OCR, Broken Line Breaks and Bad Segmentation Before You Translate

People searching translate scanned PDF, OCR translation, scanned document translation, translate PDF without losing formatting, or how to translate documents faster often assume the translation itself is the main job. In messy files, it is not. A large part of the delay can come from bad source text: broken line endings, missing characters, incorrect OCR, mixed reading order, duplicated headers, corrupted symbols, split words, invisible text layers, and tables that have collapsed into nonsense.

Fast document translation therefore begins with a question that sounds almost too simple: is the source text actually usable? If a scanned PDF has been badly recognized, every later stage inherits the damage. A translator may waste time researching a “word” that never existed, repairing sentence structure that was broken by line extraction, or translating text in the wrong order because a two-column page was flattened incorrectly.

Source cleanup is the pre-translation process of making extracted text trustworthy enough to translate efficiently. It does not mean rewriting the author. It means separating real source-language difficulty from extraction noise. When that distinction is made early, OCR translation becomes faster, more accurate and easier to verify.

Quick answer

Before translating a scanned PDF, image, old document or badly extracted file, verify that the text layer is complete and in the correct reading order. Fix obvious OCR errors, broken hyphenation, line-wrap damage, duplicated headers, missing symbols and table structure before translation. Preserve uncertain text visibly instead of guessing. Then translate from the cleaned source while keeping the original page available for comparison.

Why bad source text creates fake translation problems

Suppose a scanned manual contains the sentence:

“Disconnect the power supply before removing the protective cover.”

Poor OCR produces:

“Disconnecl the power supp1y be- fore removing the proteclive c0ver.”

A human can probably reconstruct the meaning. A translation system may also recover much of it. But every corrupted token introduces unnecessary uncertainty.

Now imagine the same problem across 120 pages.

The translator is no longer solving only translation problems. The translator is solving:

• character-recognition problems;
• word-boundary problems;
• reading-order problems;
• formatting problems;
• source reconstruction problems;
• translation problems.

If these layers are mixed together, work slows dramatically.

The first productivity principle is therefore:

repair the input once instead of compensating for the same input damage repeatedly.

OCR is not translation

Optical character recognition, or OCR, converts visible text in an image into machine-readable characters.

Translation converts meaning from one language into another.

The two processes are related but different.

A scanned page may require:

image → OCR text → cleaned source text → translation → target review → document reconstruction.

If the OCR stage is wrong, the translation stage receives the wrong source.

This is why a polished translated PDF can still be unreliable. Good formatting does not prove that the underlying text was recognized correctly.

The three source states

Before translating a document, classify it into one of three states.

State 1: clean native text

The document contains selectable text in the correct order. Copying a paragraph produces sensible text. Search works. Symbols and punctuation appear correctly.

This is the easiest state.

State 2: usable OCR with local errors

The document is scanned or image-based, but OCR produces mostly correct text. A few names, symbols, line breaks or unusual fonts require correction.

This is workable with a cleanup pass.

State 3: unreliable extraction

The OCR misses lines, merges columns, invents characters, loses table structure, or confuses large sections.

This state requires source recovery before serious translation.

Trying to translate State 3 as if it were State 1 is where large amounts of time disappear.

A 60-second source quality test

Before starting, test one ordinary paragraph and one difficult area.

Test 1: select text

Can you highlight a sentence as text rather than as an image?

Test 2: copy and paste

Paste it into a plain-text editor.

Check whether:

• words remain intact;
• punctuation survives;
• line order is sensible;
• columns have not interleaved;
• special characters remain recognizable.

Test 3: search

Search for a visible word from the page.

If search cannot find it, the text layer may be absent or unreliable.

Test 4: inspect a hard case

Find a page with:

• a table;
• small print;
• italics;
• superscripts;
• footnotes;
• multiple columns;
• a stamp or handwritten note.

A document that handles ordinary paragraphs well may still fail badly on these structures.

The source-cleanup pipeline

A robust workflow is:

inspect → extract → compare → normalize → mark uncertainty → segment → translate → verify against page.

Each step has one job.

Inspect

Determine whether the file is native text, scan, photo, mixed document or damaged export.

Extract

Use the available text layer or OCR.

Compare

Check extracted text against the visible page.

Normalize

Fix mechanical corruption without changing authorial meaning.

Mark uncertainty

Flag anything you cannot reconstruct confidently.

Segment

Create usable sentence or paragraph units.

Translate

Work from the cleaned source.

Verify against page

Return to the original visual document for high-risk details.

What should be cleaned before translation?

The rule is conservative: fix mechanical damage, not content.

Good cleanup targets include:

• split words caused by line endings;
• accidental line breaks inside sentences;
• duplicated page headers;
• duplicated footers;
• obvious OCR character substitutions;
• missing spaces;
• false spaces inside words;
• repeated lines created during extraction;
• incorrect reading order;
• broken bullet markers;
• table rows that can be reconstructed confidently;
• quotation marks converted into random symbols;
• common ligature errors;
• page numbers inserted inside sentences.

Do not silently “improve” unclear source writing. If the author’s sentence is genuinely ambiguous, preserve the ambiguity until it is solved through context or evidence.

Worked example 1: broken line-end hyphenation

Visible page:

“The system automatically re- starts after the temperature returns to the normal range.”

The hyphen belongs to layout, not to the word.

Clean source:

“The system automatically restarts after the temperature returns to the normal range.”

If the translator leaves “re- starts”, a lookup tool may treat it as two items. A machine translator may still cope, but the noise is unnecessary.

The repair is mechanical and safe because the word is visually split by the line break.

Worked example 2: real hyphen versus line-wrap hyphen

Visible page:

“This is a high- risk procedure.”

Here the correct form is “high-risk”.

Removing the hyphen completely would be wrong.

Cleanup therefore requires language awareness.

The question is not “Is there a hyphen at the end of the line?”

The question is “Does the hyphen belong to the word or only to the line break?”

Automated cleanup rules must be checked for exactly this reason.

Worked example 3: character confusion

OCR output:

“The valve opens when pressure exceeds l00 kPa.”

The first character in “100” is a lowercase L.

This is dangerous because the string is still visually plausible.

A translator may notice it. A system may not.

High-risk character substitutions often involve:

• O and 0;
• I, l and 1;
• S and 5;
• B and 8;
• rn and m;
• cl and d;
• punctuation confused with marks or dust.

Numerical and technical data deserves direct visual checking.

Worked example 4: reading order in a two-column page

A page contains two columns.

The visible reading order is:

left column paragraph 1 → left column paragraph 2 → right column paragraph 1 → right column paragraph 2.

Bad extraction produces:

left line 1 → right line 1 → left line 2 → right line 2.

The words are all present, but the sentences are destroyed.

This is not a translation problem. It is a reading-order problem.

The fastest response is to fix extraction or reconstruct columns before translating. Trying to infer meaning from interleaved lines wastes cognitive effort and increases error risk.

Worked example 5: tables

Tables create a special source-cleanup challenge because position carries meaning.

A row may connect:

• a parameter;
• a value;
• a unit;
• a limit;
• a note.

If OCR flattens the row into one long line, the translator can lose the relationships.

For example, the visible table might show:

Operating temperature | 5–40 | °C | indoor use

A poor extraction could produce:

“Operating temperature 5 40 indoor use °C”

The lexical items are present but the structure is ambiguous.

Do not translate the flattened string as ordinary prose. Reconstruct the row first.

Worked example 6: footnotes and superscripts

Academic and legal documents often use superscript numbers or symbols.

OCR may attach them to the wrong word or convert them into ordinary digits.

Source:

“The samples were stored at −20°C.¹”

OCR:

“The samples were stored at -20°C1.”

The translator must know whether “1” is part of the temperature, a footnote marker or a recognition error.

Visual comparison resolves the problem quickly.

This is why the original page should remain open during translation.

Clean line breaks before sentence segmentation

Many translation systems work segment by segment.

Bad line breaks create bad segments.

Suppose a PDF extracts:

“The committee reviewed the proposal and approved the revised budget after the additional evidence was submitted.”

If each visual line becomes a separate segment, the translator sees three fragments instead of one sentence.

That damages context and increases repeated reading.

Normalize paragraph line wraps before segmentation whenever possible.

The goal is to make segments reflect linguistic units rather than page geometry.

Do not remove meaningful paragraph boundaries

The opposite error also occurs.

Over-aggressive cleanup merges everything into one giant block.

Paragraph boundaries may signal:

• topic shifts;
• speaker changes;
• list structure;
• procedural steps;
• legal clauses;
• separate warnings;
• headings and subheadings.

Clean source text should preserve meaningful structure even while removing accidental line noise.

Headers and footers: small text, large waste

Page headers and footers are often repeated dozens of times.

Examples:

• document title;
• company name;
• confidentiality notice;
• page number;
• chapter title;
• copyright line.

If they are included in the translation stream on every page, the translator may repeatedly process material that should be handled once.

Before translating, identify recurring page furniture.

Then choose deliberately:

• translate once and reuse;
• exclude if it is not part of the required content;
• preserve unchanged if instructed;
• handle during final document reconstruction.

Repeated noise is a classic source of avoidable time loss.

Preserve uncertainty instead of inventing certainty

A damaged scan may contain a word you genuinely cannot read.

Do not silently guess.

Use a visible uncertainty marker in the working source, such as:

[unclear]

[illegible word]

[uncertain: pressure/process?]

or another project-approved convention.

The purpose is to keep the uncertainty attached to the right location.

This allows later resolution through:

• a better scan;
• another page;
• repeated terminology;
• client clarification;
• domain context.

Guessing feels fast because it closes the problem immediately. It is often slow because the guess contaminates later translation decisions.

The source truth hierarchy

When different representations disagree, use a hierarchy.

A practical order is:

1. visible original page;
2. higher-quality original file if available;
3. verified source text layer;
4. OCR output;
5. translator inference.

The exact hierarchy may vary by project, but the principle is stable: generated text should not outrank the document it was generated from.

Mixed-language pages

Some documents contain more than one source language.

Examples include:

• bilingual forms;
• product names in English inside another language;
• Latin scientific names;
• quotations;
• multilingual warnings;
• signatures and stamps;
• transliterated names.

OCR may mis-detect these regions and apply the wrong character model.

Before translation, mark which text should:

• be translated;
• remain unchanged;
• be transliterated;
• be verified as a name;
• be treated as a quotation.

This prevents a system from “translating” content that is already in the target language or from corrupting proper nouns.

Source cleanup and document privacy

OCR tools may process files locally or upload them to remote servers.

For confidential documents, check:

• where the file is processed;
• whether it is stored;
• retention period;
• whether uploaded content is used for model training;
• whether deletion is available;
• whether the service is approved for the project.

A convenient OCR service is not automatically an appropriate OCR service.

The source-cleanup stage is part of the information-security chain.

Source cleanup for AI translation

Large language models and machine-translation systems can sometimes infer through noisy OCR surprisingly well.

That is useful but dangerous.

A model may “repair” a corrupted source by predicting what probably belongs there. The prediction can be fluent and wrong.

For low-stakes gist reading, this may be acceptable.

For important translation, distinguish:

• text visibly present in the source;
• text confidently reconstructed from layout;
• text inferred by a model.

Do not let probabilistic repair silently become source truth.

A minimal cleanup pass

For an ordinary scanned document with decent OCR, use this short sequence:

1. remove repeated headers and footers from the working translation stream;
2. join line-wrapped sentences;
3. repair obvious split words;
4. check numbers and units visually;
5. check names and abbreviations;
6. inspect tables and multi-column pages;
7. mark unclear text;
8. begin translation.

This can take minutes and save much more later.

A deeper cleanup pass for difficult archives

Older scans may need more care.

Inspect:

• skewed pages;
• faded ink;
• bleed-through from the reverse side;
• unusual fonts;
• archaic spelling;
• handwritten annotations;
• stamps;
• marginalia;
• damaged paper;
• missing page edges;
• page sequence.

In such cases, translation should not outrun source recovery.

If the source cannot be read reliably, the correct professional action may be to state the limitation rather than manufacture a clean translation from uncertain evidence.

Failure mode 1: translating directly from a bad copy-paste

The translator copies a PDF into a document and immediately begins translating.

Halfway through, they discover that every page header has been inserted into the middle of sentences and that some paragraphs are out of order.

Repair:

Perform the 60-second source quality test first.

Failure mode 2: trusting the OCR because most of it looks good

OCR can be 99 percent readable and still fail on the one percent containing names, numbers or unusual symbols.

Repair:

Use targeted verification for high-risk details.

Failure mode 3: cleaning beyond the evidence

The translator encounters awkward source wording and “fixes” it into what the author probably meant.

Repair:

Separate mechanical cleanup from editorial correction. Preserve genuine ambiguity and source defects unless the project explicitly authorizes source editing.

Failure mode 4: losing the original page

The translator exports OCR text and closes the scan.

Later, a table or superscript becomes impossible to reconstruct.

Repair:

Keep source image and cleaned text side by side.

Failure mode 5: translating page furniture repeatedly

The same header is translated fifty times.

Repair:

Identify recurring non-body content before drafting.

Failure mode 6: flattening layout-dependent meaning

Forms, tables and diagrams are treated like paragraphs.

Repair:

Ask what relationships are carried by position, boxes, arrows, rows or columns. Preserve those relationships before translating the words.

Failure mode 7: using search-and-replace without boundaries

A cleanup rule replaces every “l” with “1” because several numbers were misrecognized.

The document becomes corrupted in a new way.

Repair:

Automate only patterns you can define safely. Review replacements in context.

The clean-source principle for CAT tools

CAT tools become more useful when segmentation is clean.

A damaged source can create:

• fragmented translation memory entries;
• false repetitions;
• inconsistent sentence boundaries;
• broken tags;
• useless matches in future projects.

Cleaning before import improves not only the current translation but also the reusable assets produced from it.

This is a long-term speed advantage.

Source cleanup and translation memory

Imagine the source sentence:

“Press the reset button for three seconds.”

Bad OCR stores it as:

“Press the reset butt0n for three sec0nds.”

If that corrupted segment enters translation memory, later matching becomes weaker and the error may persist.

Clean source text protects memory quality.

The same principle applies to terminology extraction and corpus building: poor input creates poor reuse.

How to handle OCR uncertainty efficiently

Use three confidence classes.

Clear

The visible page and OCR agree.

Translate normally.

Recoverable

OCR is wrong but the visible page is readable.

Correct the source and continue.

Unresolved

Neither OCR nor visual inspection gives a reliable reading.

Mark it and escalate or return later.

This prevents an unresolved word from blocking an otherwise usable page.

The batch-repair advantage

OCR errors often repeat because the same font creates the same recognition mistake.

For example, a particular scan may repeatedly recognize “rn” as “m”.

Once the pattern is identified, search the extracted text for similar cases.

Do not wait to discover the same error twenty times during translation.

Batch repair is one of the clearest ways source cleanup creates speed.

Use page anchors during long-document translation

When working with long documents, preserve page references in the cleaned source.

For example:

[Page 17]

[Page 18]

These anchors make it easy to return to the original visual page when verifying:

• a table;
• a footnote;
• a missing word;
• a figure caption;
• a number.

Without anchors, the translator may spend more time locating evidence than solving the translation itself.

Source cleanup for screenshots and mobile photos

Photos create additional problems:

• perspective distortion;
• glare;
• shadows;
• curved pages;
• cropped text;
• low resolution;
• mixed backgrounds.

Before OCR, improve the image if possible:

• crop to the relevant page;
• straighten perspective;
• increase legibility;
• avoid shadows;
• capture the whole text region;
• retake the image rather than trying to rescue unreadable pixels.

A better input image can save more time than a more sophisticated translation prompt.

Source cleanup for subtitles and transcripts

The same principle extends beyond PDFs.

A transcript may contain:

• wrong speaker labels;
• misheard names;
• missing punctuation;
• broken sentence boundaries;
• duplicated speech;
• timestamps inserted inside words.

Before translation, decide which elements are source content and which are transcription artifacts.

Again, the translator should not be forced to solve recognition and translation simultaneously when the stages can be separated.

Source cleanup for websites and copied text

Web text can carry hidden noise:

• navigation labels;
• cookie banners;
• repeated menus;
• alt text;
• code fragments;
• hidden accessibility content;
• duplicated mobile and desktop elements.

If the goal is to translate the article body, isolate the body first.

More text is not always more context. Sometimes it is just more interference.

A practical verification pass after translation

When the target draft is complete, return to the original page rather than only to the cleaned source.

Verify:

• headings;
• page labels;
• numbers;
• units;
• names;
• footnotes;
• table relationships;
• figure labels;
• warning symbols;
• checkboxes and form fields;
• omitted marginal text that matters.

This catches cleanup mistakes as well as translation mistakes.

How source cleanup improves speed without lowering standards

The speed gain does not come from skipping verification.

It comes from moving verification to the correct layer.

Instead of repeatedly wondering whether strange source text is a linguistic difficulty or an OCR error, you establish source reliability early.

Instead of fixing the same line-break problem sentence by sentence, you normalize it once.

Instead of translating duplicate headers fifty times, you identify them once.

Instead of discovering table damage halfway through, you inspect layout before drafting.

That is process compression.

Transfer: clean the problem before solving the problem

The underlying habit appears in many fields.

A scientist checks whether an instrument is calibrated before interpreting measurements.

A programmer verifies input data before debugging an algorithm.

A student checks whether they copied the question correctly before solving it.

A translator checks whether the source text is real, complete and ordered before translating it.

The principle is simple:

do not spend high-level reasoning on low-level corruption if the corruption can be removed first.

A complete source-cleanup checklist

Before translation, ask:

File state

• Is the text selectable?
• Is the file native text, scanned image or mixed?
• Is OCR required?

Completeness

• Are all pages present?
• Are any edges cropped?
• Are any lines missing?

Reading order

• Are columns extracted correctly?
• Are footnotes in the right place?
• Are tables still meaningful?

Character quality

• Are names readable?
• Are numbers correct?
• Are symbols preserved?
• Are special characters corrupted?

Line structure

• Are words split at line endings?
• Are sentences broken into artificial segments?
• Are paragraph boundaries preserved?

Repetition

• Are headers, footers or page numbers polluting the body text?

Uncertainty

• Are unclear areas visibly marked rather than guessed?

Security

• Is the OCR method appropriate for the document’s confidentiality level?

If these are controlled, translation begins from a much stronger base.

The central idea

People translate quickly when they stop translating noise.

A scanned document can contain two different difficulties:

1. the language itself;
2. damage introduced by capture and extraction.

Those difficulties should not be solved at the same time if they can be separated.

The source-cleanup method is:

inspect → recover → normalize → mark uncertainty → segment → translate → verify against the original page.

The cleaner the source, the more of the translator’s attention can be spent on meaning rather than repair.

Summary

Source cleanup is the process of preparing extracted or OCR-recognized text before translation. It removes mechanical noise such as broken line wraps, split words, duplicated headers, obvious character-recognition errors and bad reading order while preserving genuine ambiguity and authorial wording.

For scanned PDF translation, OCR translation and image-based document translation, this step can save substantial time because it prevents repeated low-level repairs during drafting. It also improves segmentation, terminology consistency, translation-memory quality and final verification.

Fast document translation does not begin with faster target-language output.

It begins with a source that deserves to be trusted.

Frequently asked questions

What is OCR translation?

OCR translation is a workflow in which text is first recognized from an image or scanned document and then translated into another language.

Why can’t some scanned PDFs be translated directly?

Because they contain page images rather than machine-readable text. OCR is needed to recover characters before translation.

Should OCR text be corrected before translation?

Yes when the errors are mechanical and clear. Obvious character substitutions, broken line endings and reading-order problems should be repaired before they create translation errors.

Should a translator correct bad source writing during cleanup?

Not automatically. Cleanup should fix extraction damage, not silently rewrite genuine source ambiguity or poor writing unless source editing is explicitly part of the job.

How do I know whether a PDF has a text layer?

Try selecting and copying a sentence, then search for a visible word. If selection and search fail, the page may be image-only.

What OCR errors are most dangerous?

Errors involving names, numbers, units, negation, technical terms, table relationships and symbols can create serious meaning changes.

Can AI fix OCR errors automatically?

AI can often infer likely text, but inferred reconstruction should not silently replace evidence from the original page in high-stakes work. Verify important repairs visually.

Does source cleanup help CAT tools?

Yes. Clean segmentation and reliable source text improve translation-memory matches, terminology extraction and future reuse.

Internal-link opportunities

This article can connect naturally to these existing eduKateSG translation owners:

• How People Translate Quickly | Translation Triage: Sort Easy, Risky and Research-Heavy Segments Before You Start — for deciding which damaged areas deserve immediate attention.
• How People Translate Quickly | Sentence Skeletons: Find the Main Clause Before You Translate — after extraction noise has been removed from long sentences.
• How People Translate Quickly | Context Prediction: Narrow the Meaning Before You Look Up a Word — for genuine ambiguity that remains after OCR repair.
• How People Translate Quickly | String Context Packs: Use Screenshots, Developer Notes and Character Limits to Translate UI Faster — for cases where visible layout supplies missing context.
• How People Translate Quickly | Pattern Reuse: Use Collocations, Glossaries and Translation Memory — for protecting reusable assets from corrupted source text.
• Translate Easily to any Language | Translate Text, Speech, Images and Documents Without Losing Meaning — for the broader document-translation route.
• Translate Precisely | Names, Numbers, Dates and Units — The Details That Cannot Drift — for high-risk details that should be checked against the original page.

Measure the cleanup so you know whether it is saving time

A cleanup stage should not become a hobby. Its purpose is to reduce downstream translation friction. That means the translator should be able to tell whether the repair was worth doing.

A simple way to measure this is to compare a short untreated sample with a cleaned sample. Take two passages of similar difficulty. Translate the first directly from the raw extraction. For the second, spend a few minutes fixing obvious OCR errors, broken paragraphs, headers, reading order and segmentation before translating. Record four things: total time, number of lookups, number of interruptions caused by source defects and number of corrections discovered during final verification. The goal is not to make the source beautiful. The goal is to reduce avoidable translation work.

Suppose a raw ten-page scan produces thirty apparent unknown words. After checking the page image, you discover that twelve are OCR substitutions, six are broken compounds and four are fragments from a footer. Only eight are genuine lexical problems. Cleaning those defects before translation has not merely improved typography. It has changed the problem from thirty research decisions to eight. That is a major speed gain because the translator is no longer spending linguistic attention on errors created by the extraction process.

The same principle applies to segmentation. If a PDF converter breaks one paragraph into seven text boxes, a CAT tool may present seven artificial segments. The translator then has to reconstruct the argument repeatedly. Repairing the paragraph boundary once can save multiple interpretation cycles. Good source cleanup therefore creates leverage: one upstream correction prevents many downstream micro-corrections.

Know when OCR is the wrong first move

OCR is useful when the source is an image of text, but it is not automatically the best route for every difficult document. A native PDF may already contain a reliable text layer even when its visual formatting is complicated. Running OCR over that file can replace correct characters with uncertain ones. A bilingual PDF may contain selectable text in one layer and scanned annotations in another. A table may be easier to transcribe structurally than to force through generic OCR. Handwriting may need human reading rather than confident-looking machine output.

Before starting OCR, ask a narrower question: what part of this document is actually inaccessible to the translation workflow? If only three image captions are scanned, OCR those captions rather than rebuilding the entire document. If the body text is selectable but the reading order is wrong, extraction and reordering may be enough. If the document is a photographed form with checkboxes, stamps and marginal notes, preserve the image and create a controlled transcription beside it instead of pretending the entire page is ordinary running text.

This selective approach protects both speed and source integrity. The fastest translator is not the person who applies the most automation. It is the person who applies the smallest reliable repair that turns the source into a trustworthy translation input.

Keep a source-integrity log for uncertain repairs

When cleanup changes the source representation, important uncertainty should remain visible. A compact source-integrity log can record page number, suspicious text, what the image appears to show, the normalized form used for translation and the confidence level of that repair.

For example:

• Page 4: OCR reads 1O mg; image appears to show 10 mg; normalized to 10 mg; high confidence after visual check.
• Page 7: surname appears as Ramires or Ramirez; no independent confirmation; preserve uncertainty for client query.
• Page 11: table heading split across two columns; reordered according to visual alignment; high confidence.
• Page 15: handwritten annotation partly illegible; not silently reconstructed; flagged for review.

This log prevents a dangerous form of speed: quietly converting uncertain source material into certain-looking target text. It also makes collaboration faster because a reviewer can see which decisions came from language translation and which came from source reconstruction.

Separate text recovery from layout reconstruction

Another common mistake is trying to solve text extraction, translation and final page design at the same moment. These are related tasks, but they have different failure modes.

First recover trustworthy text. Then translate the meaning. Then rebuild the target document layout. Keeping these stages separate makes errors easier to locate. If a number is wrong, you can ask whether the source extraction was wrong, whether the translation changed it or whether the layout stage placed it in the wrong cell. When all three jobs are mixed together, diagnosis becomes slow because every visible defect has several possible causes.

This separation is especially useful for forms, brochures, manuals, certificates and scanned reports. A visually faithful target document depends on accurate text, but visual fidelity should not pressure the translator into preserving bad source segmentation. The target can eventually look like the original while the working translation file uses cleaner paragraphs, stable sentence boundaries and explicit table structure.

The speed lesson is simple: repair the representation before you accelerate the translation. A clean source is not cosmetic preparation. It is part of translation engineering because it determines which problems the translator must solve and which problems can be removed before language work begins.