Our own document translation technology
DocTranslator AI Engine
The DocTranslator AI Engine is the document translation technology we develop ourselves. Where most translation tools work with plain text, this one works with the page: it finds where every line sits, translates it and writes the translation back into the same place, so the file you download has the layout of the file you uploaded.
It is designed and written by the DocTranslator team.
How it works
Four stages from upload to finished file
- 01
Read
Read the file
The engine works out what the file is from its contents rather than its name: a PDF with a real text layer, a scan, a photo or an Office document. Each kind takes a different path.
- 02
Separate
Separate the words from the page
Only the text is lifted out. Images, diagrams, charts, photos, signatures and stamps stay where they are and are never sent for translation.
- 03
Translate
Translate and check
Each passage is translated and checked before it is accepted. A passage that fails a check is translated again.
- 04
Rebuild
Rebuild the page
The translation goes back into the original positions, set in fonts that contain the characters of the target language.
A · Digital PDFs
PDFs with a text layer
In a PDF exported from Word, InDesign or a similar program, the engine reads the text block by block and table cell by cell. It joins the lines of a paragraph into one passage before translating, so a sentence that wraps across three lines is translated as one sentence rather than as three fragments. It then removes the original words from the page, leaves images, diagrams and vector drawings untouched, and writes the translation into the same box.
A translation is rarely the same length as its source. German and Russian usually run longer than English, and Chinese and Japanese take fewer characters. The engine lets the text wrap inside its original block and reduces the font size only when the translation still does not fit. Table cells are handled one at a time, so figures stay in their own rows and columns.
B · Scans and photos
Scans and photos
A scanned page is a picture, so there is no text in it to replace. The engine keeps the original image as the background of the page, recognises the printed text with OCR, and then removes only the ink of those letters, filling each spot with the colour of the paper around it. Anything that is not running text is left alone. Stamps, seals, signatures, photos, charts and diagrams stay exactly as they were scanned, which is what you want on a certificate or a signed contract.
Tables keep their printed grid. The engine finds the ruled lines of the original table in the image, clears the text inside each cell without touching the borders, and writes the translation into the same cell. A line printed sideways along a margin is written back sideways.
Sometimes a line cannot be erased cleanly, for instance where the text sits on a busy background. In that case the engine leaves the original print in place rather than smearing it, and adds an invisible copy of the text so the page can still be searched.
C · Structured files
Word, Excel, PowerPoint, EPUB and HTML
Office files already keep their text apart from their layout, so the engine translates them from the inside. It replaces the words of each paragraph, slide and spreadsheet cell and keeps bold, italic and other formatting on the words it belonged to, even when the translation puts those words in a different order.
EPUB books and HTML pages are translated text by text, and their markup is left as it was, so chapters, headings and links keep working in the translated copy.
ENThe report is ready today.
↓
DEDer Bericht ist heute fertig.The bold word moves to the end of the German sentence, and its formatting moves with it.
Quality
Checks before a file comes back
AI translation can fail in ways that are easy to miss on a quick read, so the engine checks each passage before it accepts it:
rejectAn answer that repeats the original instead of translating it is rejected.rejectSo is an answer in the wrong language, one that loops or stops halfway, and one with stray characters from another script.retryA rejected passage is translated again, first together with the lines around it for context, then on its own.compareNumbers, web addresses and email addresses are compared between the original and the translation, so a changed figure is flagged.reuseA sentence that appears on several pages is translated once, so a heading or a term reads the same everywhere in the document.discardIf a large part of a document comes back untranslated, the engine throws that result away rather than hand over a half-translated file.
Built and run by DocTranslator
Our own code, from reading the page to rebuilding it.
The engine is developed by the DocTranslator team. The page analysis, the handling of scans and tables, the rebuilding of each page and the checks above are our own code. Because we write the engine ourselves, a problem such as a shifted table or a missing character can be fixed in the engine itself.
Translations into Chinese, Japanese, Korean, Hindi, Thai and other scripts are set in fonts that contain those characters, so you do not get empty boxes where letters should be. Only the characters a document actually uses are embedded, which keeps the file size down.
DocTranslator AI Engine FAQ
Is the DocTranslator AI Engine built by DocTranslator?
Yes. The engine is developed in-house by the DocTranslator team. The page analysis, scan handling, layout reconstruction and quality checks are our own code.
Does it keep stamps, seals and signatures on scanned documents?
Yes. The engine removes only the ink of recognised running text. Stamps, seals, signatures, photos, charts and diagrams are left as they were scanned, so a certificate still shows its seal and a contract its signatures after translation.
What happens to tables?
In a PDF with a text layer, each cell is translated on its own and written back into the same cell. On a scan, the engine finds the printed grid in the image, clears the text inside each cell without touching the lines, and writes the translation into the cell it came from.
Will the translated file look exactly like the original?
Close, though not always identical. A translation is usually longer or shorter than its source, so text can wrap differently or be set slightly smaller to fit its box. On a dense page, such as a photographed table, take a quick look at the result before you send the file on.
Which languages does DocTranslator support?
DocTranslator supports 108 languages, and you can translate between any two of them. The full language list shows every one.
Can I use it for a visa or court filing?
The engine produces AI translations. Immigration offices, courts and universities usually ask for a certified translation, signed by a professional translator who confirms it is accurate and complete. For that, use our certified translation service.
Translate a document and check the layout yourself
Upload a PDF, a scan or an Office file, choose two of 108 languages and compare the result with your original. Translation starts at $0.99 per document.
Related Tools
File Formats
Document Types