Back-translation: a simple way to spot errors
How translating a translation back into your own language can show up errors, and what the test can't tell you.
Back-translation is a round trip. You translate your text into the other language, then translate that result back into yours, and compare. If the meaning drifts, something probably changed on the way.
1 A back-translation is a translation of a translated text back into the original language. It's "made without reference to the original text." In machine translation it's also called a round-trip translation.
How to run the test
- Translate your original into the other language and save the result.
- Copy only the translated text. Don't copy your original along with it.
- Paste it into a fresh window, ideally a different tool, and translate it back.
- Put your original and the returned text side by side.
- Mark every difference in names, numbers, dates and meaning.
A different tool for the return trip gives you a second opinion, though a fresh window of the same tool also works.
What the test shows
A difference points you at a spot to look at. Typical flags are a missing sentence, a changed number, a word that now means something else, or a request that has turned into an order.
It also helps when you can't read the language at all. It's one of the few checks you can run alone.
What it misses
The same reference is careful about this. Comparing a back-translation with the original1 is "sometimes used as a check on the accuracy of the original translation". It adds that the results "while useful as approximate checks, are not always precisely reliable". The reason it gives is that1 "linguistic symbols ( words ) are often ambiguous".
In practice that means two things.
- A clean round trip can still hide an error.2 These tools "tend to produce fairly literal translations", so a stiff or impolite phrase in the other language may come back as your original wording and look fine.
- A messy round trip can come from the return trip. The mistake may be in the second translation, not the first.
So treat a match as a good sign, not as proof. Treat a mismatch as a question to chase.
| If the round trip shows | Likely next step |
|---|---|
| Same meaning, same numbers | Still get a reader of the language for anything important |
| A changed number or name | Fix it in the original wording and translate again |
| A different meaning | Rewrite the sentence in simpler words and test again |
| A missing sentence | Translate that part on its own |
The next steps are common sense, not findings from a source.3 Research says neural output typically needs a human editor, and a back-translation doesn't replace that person. For the full review, see what to check before sending.
Where the facts come from
- Translation, Wikipedia
- Neural machine translation, Wikipedia
- Machine translation, Wikipedia
Keep reading
How AI document translation works and where it goes wrongWhat an AI translation tool is actually doing with your text, and the kinds of mistakes it makes with business documents.
AI translation compared with a human translatorWhen a machine translation is good enough for a business document, and when a trained human translator needs to do the work.
What to check before sending a translated documentA practical way to review a machine translation when you don't speak the other language, from names and numbers to who reads it.