How to translate a long document with ChatGPT
Working with documentsLast checked
Short answer
Work section by section, and keep a glossary you re-paste with each one. Translation output is about as long as the input, so you need room for both, which means documents that fit comfortably for summarising do not fit for translating. Consistency of terminology is the thing that separates a good result from an obviously machine-made one.
Translation is the hardest case for length
Most document tasks produce output much shorter than the input. A summary of fifty pages is one page.
Translation produces output roughly as long as the input, sometimes longer depending on the language pair. So you need capacity for the document and its translation at once.
This is why a document you have happily summarised may fail as a translation. It is not a different limit, it is twice the demand on the same one.
Truncation is worse here than anywhere else
A truncated summary is incomplete. A truncated translation is a document that stops mid-thought, and if you do not read the target language well, you may not notice which part is missing. Always check the last section came through.
The glossary is what makes it good
The single biggest quality difference, and it costs almost nothing.
Long documents use the same terms repeatedly. Translated section by section without a shared reference, the same term comes out three different ways, and the result reads as machine-made even when each individual sentence is fine.
Translate the first section, and ask for the terms
"Translate this into French. Then list every technical or recurring term you translated, with the translation you chose, as a glossary."
Review the glossary yourself
This is where your judgement is worth most. You know which term is the company's standard usage and which is a reasonable but wrong choice.
Re-paste the glossary with every later section
"Use this glossary for consistency. Do not deviate from it." Then the next section.
Add to it as you go
New terms appear. Keep the list growing and keep sending it.
Work in markdown
A translation that arrives as one undifferentiated block, when the original had headings, lists and tables, means rebuilding the structure by hand afterwards.
Converting to markdown first and asking for markdown out keeps all of that intact. Headings stay headings, lists stay lists, tables stay tables, and the structure survives the round trip.
Instructions worth giving
Set the register. "Formal business register" or "plain language for a general reader" changes the output substantially, and left unsaid you get an average of everything.
Say what not to translate. Product names, brand names, legal entity names, code. Otherwise they get helpfully translated and you have to find them all again.
Ask about ambiguity. "Where the source is ambiguous, translate it and flag the ambiguity in a note." Genuinely useful, because it surfaces places the original was unclear, which is worth knowing independent of the translation.
Ask for the untranslatable. "Note anything that does not carry across cleanly and explain why." Idioms, culturally specific references, legal terms without an equivalent. These are exactly where a silent translation goes wrong.
Where to be careful
Anything contractual or regulated. Have a qualified translator review it. A mistranslated obligation is expensive and does not look like an error, it looks like a sentence.
Languages with less training data. Quality varies considerably by language pair, and confidence does not.
Documents where precision is the point. Medical, legal, safety. Understanding a document is one thing. Relying on the translation is another.
Scanned originals. OCR first, and check it, because OCR errors become translation errors that read as perfectly fluent sentences meaning the wrong thing.
Common questions
Can it translate a whole document in one go?
Only a short one. Translation output is roughly as long as the input, so you need room for both, and a document that fits for summarising may not fit for translating. Section by section is the reliable approach.
How do I keep terminology consistent across sections?
Build a glossary as you go and re-paste it with each section. Without it, the same term gets translated three different ways across a long document, which is the most common flaw in machine translated documents.
Will formatting survive?
Not reliably from a PDF or Word upload. Working in markdown preserves headings, lists and tables through the translation, which saves rebuilding the structure afterwards.
Is it good enough for professional use?
For understanding a document, usually yes. For anything published, contractual or regulated, have a qualified human review it. Translation errors in a contract are expensive and they do not look like errors.
Keep reading
How to split a long document for ChatGPT
Keep each part under 10,000 characters, cut at section boundaries, and send the wait-for-all-parts instruction first. Past 10,000 a paste becomes an upload.
The 30 second check for whether ChatGPT read your whole document
Files past 2 million tokens are truncated with no warning. One question about the final page tells you how much ChatGPT really read, in 30 seconds.
How to summarise long documents with ChatGPT
Anything past 2 million tokens is cut off silently, so run the last section test before you read the summary. Then the prompts that stop it coming back bland.