You’re three hours into reformatting a 40-page financial report that came back from machine translation. Tables are scrambled, bullet points are orphaned across pages, fonts are mangled, and that carefully designed contract template? It looks like it survived a printer jam. Welcome to the formatting graveyard—where machine translation and document structure go to die.
The problem isn’t the translation itself. It’s that most MT systems strip formatting codes before they translate, work on raw text, and then hand back a pile of words with zero structural information. The layout reconstruction that follows? It’s either nonexistent or crude enough that you spend more time fixing format than refining language.
This isn’t an edge case—it’s the norm. According to a 2025 translator survey, 66% report that MT output requires significant edits after post-processing, and a quarter say the quality is simply poor. Yet 87% of professionals still use MTPE (machine translation post-editing) because, when it works, it can save up to 63% of time on repetitive work. The trick is knowing when to use which method, and how to structure your workflow to catch formatting breakage before it costs you hours.
Let’s walk through why MT destroys formatting, what actually works, and which approach fits your situation.
The Technical Reason MT Mangles Your Documents¶
Machine translation systems work on text, not documents. This fundamental difference explains everything that goes wrong.
Here’s what happens:
1. Formatting codes are stripped during extraction. When MT reads your DOCX or PDF, it doesn’t see “Heading 1 in Calibri Bold.” It sees raw text: “Company Financial Summary.” All the metadata—font, size, color, paragraph styling—is dropped. Same with tables: the structural markup goes away, leaving just the cell contents.
2. The text is translated in isolation. The MT engine translates that plain text without any awareness of how it fits into a larger layout. A sentence that’s 14 words in English might become 18 words in German or 10 words in Chinese. The system doesn’t care—it just produces the translation.
3. Formatting is reapplied blindly. The tool tries to reinsert codes back, but it’s making educated guesses. Did that period end a paragraph or a sentence? Was that line break intentional or just word wrap? The result is best-effort reconstruction, often wrong.
4. Complex structures (tables, multi-column layouts, headers tied to footers) break entirely. When table cells contain different amounts of text pre- and post-translation, the whole structure destabilizes. Images that were anchored to specific text now float. Headers that were one line span three. Tables in PDFs? Forget it—PDFs have fixed pixel-based layouts, and text can’t reflow.
This is why professionals say: PDF is your enemy for translation. The format’s rigid structure means zero flexibility for text expansion or contraction. The same document in DOCX can reflow; in PDF, it’s a brick.
Why Text Length Changes Break Everything¶
Here’s a number that matters: translated text is rarely the same length as the source.
According to translation professionals, language pairs have predictable expansion or contraction ratios:
- English → German: +15% to +20% word count (German’s long compound words take space)
- English → Spanish: +10% to +15% (Spanish uses more words for the same concept)
- English → Chinese: -20% to -30% word count (Chinese characters pack meaning densely)
- English → French: +10% to +15%
That expansion isn’t abstract. When you translate a one-page English contract into German, it becomes 1.2 pages minimum. That table with 5 columns? The 6th column gets pushed down. Headers designed for one line now need two. Bullet points that fit neatly now orphan across page breaks.
Generic tools (Google Translate, ChatGPT, DeepL’s free browser tool) have no mechanism to handle this. They extract text, translate it, and throw it back—hoping the structure survives. As research shows, it rarely does.
Three Professional Approaches That Actually Work¶
Approach 1: CAT Tools (Trados, memoQ) - For Recurring Work¶
CAT (Computer-Assisted Translation) tools are built for professionals who translate repeatedly. They work at the file format level, meaning they extract text while preserving the underlying structure codes, translate the text safely inside a sandboxed segment-by-segment environment, and then reinsert the codes into a new document using the original template.
How it works:
- You open your DOCX in Trados or memoQ.
- The tool parses the file and identifies segments: paragraphs, table cells, headers, captions. Formatting tags stay attached to each segment.
- You (or MT) translate the segments inside the tool’s editor—the segment stays anchored to its formatting context.
- The tool reconstructs the document, reinserting all tags. Tables stay tables, fonts stay fonts, page breaks survive.
Strengths: - Handles complex files (Word, PowerPoint, InDesign, Excel, HTML, XML). - Translation Memory means you don’t retranslate repeated phrases. - QA checks catch missing numbers, inconsistent terminology, broken tags. - Integrates with MT engines (Google, DeepL, custom neural models) for first-draft automation.
Weaknesses: - Steep learning curve; takes weeks to set up for a team. - License costs: €500-1000+ per year per translator. - Overkill if you translate once every six months. - Requires source files to be clean (valid table structure, no embedded images in cells, consistent styles). Garbage in = garbage out.
Best for: Translation agencies, in-house teams with high volume, recurring documents (legal templates, technical manuals, compliance reports).
Approach 2: Pre-Editing Your Source - The Quick Fix¶
Before handing anything to MT, fix the document itself. This costs time upfront but saves massive post-editing work.
Checklist before sending to MT:
- Remove embedded images and text-in-graphics. If your logo or diagram has overlaid text, the MT system will try to translate that text—it’ll end up garbled or orphaned. Extract text to callouts instead.
- Break tables into simple structures. No nested tables, no merged cells, no content that spans multiple columns. Flat tables survive MT; complex ones shatter.
- Use styles consistently. Don’t apply bold directly—use “Emphasis” style. Don’t adjust spacing manually—use paragraph styles. When styles are consistent, the tool can reapply them predictably.
- Put text in text boxes, not in shapes. Shapes (especially rotated or positioned shapes) don’t translate cleanly; their position gets lost or the text inside goes untranslated.
- Remove manual formatting. Delete direct font/size/color overrides; rely on the template. When you translate, formatting attached to styles survives; direct formatting often doesn’t.
- Simplify headers and footers. Complex header layouts (images + tables + text) often break. Use simple text headers; add complexity back in post-editing.
- Flag unusual content. Add comments or notes: “This table is pricing—preserve the alignment,” or “This image has Japanese text that needs separate translation.”
Effort: 30 minutes to 2 hours per document, depending on size. Payoff: 30-50% faster post-editing because the MT system has cleaner input.
Best for: One-off documents, reports you’re translating in-house, scanned documents you’re digitizing before translation.
Approach 3: AI Tools with Built-In Formatting Preservation - Fastest for One-Offs¶
A new category of tools (like ChatsControl) is designed specifically to translate whole documents while keeping formatting intact—no CAT tool setup, no pre-editing prep work.
How it works:
- Upload your DOCX, PDF, or PowerPoint.
- The tool runs OCR if it’s a scan, then parses the document’s structure (identifying text, tables, images, headers, styles).
- It translates the content while maintaining structural awareness—tables stay tables, fonts stay fonts.
- You get back the same file format with formatting preserved and a bilingual side-by-side view for review.
- Use the chat feature to request changes to specific passages without retranslating the entire document.
Strengths: - No configuration, no learning curve—upload and go. - Handles PDFs, scans, and formatted documents automatically. - Bilingual review built in, so you see source and translation side-by-side. - Built-in QA checks for numbers, names, omitted content. - Fast: most documents return in minutes to hours, not days. - Works for one-off translations, not just high-volume repetitive work.
Weaknesses: - No translation memory (so repeated phrases aren’t tracked for consistency across projects). - Not a full team workflow tool—no vendor management, no formal QA gates. - Doesn’t support handwritten documents or very poor-quality scans (those still need manual handling).
Best for: Freelancers, one-off professional translations, bilingual review workflows, documents where speed matters, certifications you want QA’d before delivering to clients.
Side-by-Side Comparison¶
| CAT Tools (Trados, memoQ) | Pre-Editing | AI Tools (ChatsControl) | |
|---|---|---|---|
| Setup time | Days to weeks | None | 5 minutes |
| Learning curve | Steep | None | None |
| Cost | €500-1500/year + training | Time (prep work) | Per-character pricing ($20-50 typical) |
| Best for | Agency workflows, high volume | Mixed workflows | Fast one-offs, bilingual review |
| Formatting preservation | 95%+ (if source is clean) | 80-90% (depends on post-edit) | 90%+ (AI-native) |
| Translation memory | Yes | No | No |
| Team collaboration | Yes (with features) | Limited | Limited (shared links) |
| Handles PDFs well | With OCR plugin | Difficult | Yes, by default |
| Time per 5000-word doc | 4-8 hours (translation+review) | 2-4 hours (pre-edit+MT+edit) | 1-2 hours (upload+review+refine) |
How Professional Workflows Actually Handle This¶
Real-world translation of formatted documents doesn’t rely on a single approach—it’s a combination.
Typical workflow for a legal contract (20 pages, table of contents, formatted tables, signature block):
- Pre-edit the source (30 minutes). Simplify the header/footer, flatten complex tables, add comments flagging unusual formatting.
- Run through CAT tool or AI tool (2 hours). Let Trados or ChatsControl translate while preserving structure.
- Bilingual review (1.5 hours). Read source and translation side-by-side, catch mistranslations, terminology inconsistencies, and initial formatting damage.
- Post-edit formatting (1 hour). Reflow text that didn’t fit, fix orphaned bullet points, realign table content, tweak spacing.
- QA pass (30 minutes). Run automated checks (or manual scan) for: missing numbers, changed proper names, broken table structure, page breaks in weird places.
- Deliver (5 minutes). Export to final format (PDF, signed DOCX, whatever the client needs).
Total time: 5-6 hours for 20 pages. Without formatting tools, the same document takes 12-16 hours because you’re either wrestling with a generic MT tool or rebuilding the document from scratch.
Common Pitfalls That Kill Formatting (And How to Avoid Them)¶
Pitfall 1: Translating a PDF directly without OCR or source recovery. PDFs are display-first, meaning they’re optimized for viewing, not editing. Text is anchored to pixel coordinates, images are embedded, fonts are locked. Translating a PDF without converting it to DOCX first is asking for destruction.
Fix: Always request the source DOCX or PPTX. If you only have PDF, use OCR to rebuild an editable version first (ABBYY FineReader, Acrobat Pro), then translate the DOCX.
Pitfall 2: Copying MT output into your original file. You translate in Google Translate, get text back, copy-paste it into your formatted Word doc. Fonts change, spacing explodes, tables collapse.
Fix: Never copy-paste MT into your formatted original. Either use a tool that preserves formatting (CAT or AI tool), or rebuild the document from a clean template.
Pitfall 3: Assuming CAT tools handle poor-quality sources. A CAT tool is only as good as the source file. If your original DOCX has broken tables (mismatched row counts, invalid markup), or if cells contain images, the tool can’t fix it—it just inherits the mess.
Fix: Pre-edit ruthlessly. If the source is broken, the translation will be broken.
Pitfall 4: Not accounting for text expansion when you post-edit. You get back German text that’s 20% longer. You paste it back into your one-page English layout hoping it’ll fit. It doesn’t. Now half your bullet points orphan and your footer sits in the middle of the page.
Fix: Assume expansion/contraction upfront. Reserve extra space in boxes, use styles that can reflow, or rebuild pages during post-editing rather than trying to jam new text into old containers.
Pitfall 5: Treating all language pairs equally. English-Spanish? Formatting breakage is moderate (text grows 10-15%). English-Chinese? Text shrinks 20-30%, and you might need to reduce font sizes. English-Japanese? Similar to Chinese, plus you need to check for right-to-left layout issues.
Fix: Before translating to a new language, estimate text expansion using industry benchmarks. Leave space accordingly.
Putting It All Together: Your Decision Tree¶
Translating for the first time or one-off documents? → Use an AI tool like ChatsControl. Faster, no setup.
High-volume repetitive work (100+ documents/year)? → Invest in Trados or memoQ. Translation memory pays for itself after 30-50 documents.
Mix of both? → Pre-edit in Word (30 min per doc), run through AI tool or MT engine, post-edit (1-2 hours).
Scanned documents or PDFs only? → AI tool or OCR + Trados. Generic MT won’t cut it.
Need certified/legally binding translations? → Use AI for the draft (bilingual review + QA built-in), then have a sworn translator verify and stamp it.
The bottom line: formatting destruction in MT is solvable, but it requires choosing the right tool for your workflow. A CAT tool doesn’t help if you translate five documents a year; an AI tool doesn’t scale to 500. Know your volume, know your pain points, and pick accordingly.
FAQ¶
Q: Why does Google Translate mangle my PDF but Word docs are better? A: PDFs have fixed, rigid layouts—text boxes don’t expand. Word docs (DOCX) have flexible structure, so text can reflow. PDF is format-first; Word is content-first. Always request DOCX or provide editable source if possible.
Q: Can I just copy-paste MT output back into my original file? A: No—translated text is rarely the same length. German and Dutch translations run 15-20% longer than English, Chinese runs 30% shorter. Pasting breaks text boxes, pushes images around, and ruins pagination. Rebuild from the formatted source.
Q: Do CAT tools like Trados actually preserve tables? A: Yes, but only if your source file is clean (valid table syntax, consistent styles). CAT tools maintain table structure by extracting and reinserting formatting codes. If your source has broken table markup, CAT tools can’t fix it.
Q: How much faster is MTPE than translating from scratch? A: Up to 63% faster for repetitive, technical content with good MT output. Creative, legal, and medical texts see smaller gains (20-40% faster) because they need more rework. Translation quality varies by language pair (EN→DE is better than EN→Japanese).
Q: Do I need a tool like ChatsControl if I already use Trados? A: No—Trados is a full CAT/TMS for team workflows and translation memory leverage. ChatsControl is faster for one-off documents and preserves formatting automatically without configuration. Use Trados if you manage vendor workflows; use ChatsControl for quick, bilingual-reviewed drafts.
Q: What’s the difference between pre-editing and post-editing formatting? A: Pre-editing: fix source formatting before MT (clean tables, remove embedded images, simplify styles) so MT has an easier job. Post-editing: fix MT output and reflow text into the target format. Both are needed for complex documents.