Back Translation: How to Check a Translation You Can't Read

What back translation catches and what it misses #

Back translation puts the judgment in the hands of the person who knows what the text should say. You don't need to read Japanese to notice that "run" became "check". You need to remember what you meant.

It also costs nothing and works in any translator. Machine errors come in a few typical shapes: a dropped negation, a swapped subject, an idiom translated word by word, a number or date that moved. An independent return trip rarely undoes a mistake made on the way out, so these shapes tend to show up in the back translation.

People find other uses for it too. On Hacker News, one user suggests it as a preview of what a foreign reader gets: "If you want a hint of what Google Translate does to a language, do it full circle." Another uses it at conferences held in a language they speak only well enough for everyday conversation, and writes that "the round trip translation usually helps enough that I can parse it correctly."

The weak side is documented too. Dorothée Behr studied the back translation step of the 2012 European Quality of Life Survey and compared it with checks by native speakers. In Assessing the use of back translation (2017) she concludes: "While back translation can uncover problems, it causes quite a number of false alarms, and even more importantly, many problems remain hidden."

Hidden problems. The back translation reads fine while the translation is wrong for its reader. Register is the clearest case. We translated the same thank-you once as Formal and once as Informal:

Nothing in the English tells you the casual version is too casual for your manager's manager. English has no verb forms for politeness, so that information has nowhere to go on the way back. The "run the numbers" example has the same gap: くれますか is a request between peers, and "Could you please" sounds more deferential than the Japanese. A Hacker News commenter sums up the structural losses: "English to Japanese loses plurals, Japanese to English loses most honoriffics [sic] and might need to invent a subject".

False alarms. The back translation reads badly while the translation is fine. You might rewrite a sentence a native reader would have found natural.

Shared mistakes. If both directions make the same wrong assumption, the error comes home looking like a confirmation. If you meant that she ducked and both systems read "duck" as the bird, "I saw her duck" comes back word for word with the wrong meaning inside it.

Writing for the round trip. You may start rewriting your source until the round trip looks clean. One Hacker News user describes doing that: "I kept changing the English until the English-Japanese-English round trip translation wasn't wildly wrong". Plainer sentences do translate better, which is Brislin's decentering at work. A clean round trip can still sit on top of awkward Japanese, though.

How much the gaps matter depends on what you send. For a message to a host, a question to a shop or a reply on a forum, a clean back translation is enough. At worst, your reader gets a slightly odd sentence and still understands you.

For a job application, a contract, a medical question or anything sent to someone senior, use the round trip as a first filter. Check the register separately and ask a native speaker to read the result. Our post on how to tell if an AI translation is correct covers the other checks, and how to use a translator to learn a language turns the round trip into a study routine.

How Fink uses back translation #

We built Fink for people who send translations they can't fully read and want to understand them first. Every translation comes with an Insights panel, and the back translation is one part of it:

• Back translation. A separate, smaller Mistral model writes it, working only from the translation. It never sees your original, so it can't copy your wording back. If you edit the translation, you can refresh it.
• Meaning notes explain what the translated wording conveys: tone, idioms, ambiguity and alternative readings. They cover what the round trip loses. For the thank-you, the note flagged ご協力 as a polite, honorific word for cooperation.
• Pronunciation, in Hepburn romanization for Japanese, so you can read the result aloud.

Next to the panel, a quality check based on the MQM error typology marks problems in the translation itself, including a formality level that doesn't fit. You choose Formal, Neutral or Informal before translating, and a glossary keeps names and terms consistent.

Fink's back translation has the limits described above. Both models come from the same vendor and can share a blind spot, and the too-deferential "Could you please" in the "run the numbers" example came from our own back translation.

You can try Fink as a guest without an account. The free plan includes 3 credits a month, about 150 translations, and the pricing page lists the rest.

Questions people ask about back translation #

What is back translation? Back translation means translating a translated text back into its original language and comparing the result with the source. It lets someone who reads only the source language check whether the meaning survived.

What is the difference between forward and back translation? Forward translation produces the text your reader will see. Back translation runs the other way, from the target language back into the source language, and serves only as a check.

Is back translation reliable? It catches many changes in meaning, such as a dropped negation or a literal idiom. It misses register, politeness and naturalness, and it raises false alarms. Dorothée Behr's 2017 study found that many problems stayed hidden, and she recommends having bilingual reviewers check the translation itself.

Is round-trip translation the same as back translation? Yes. Round-trip translation, reverse translation and back translation name the same procedure. Machine translation researchers tend to say round-trip translation, while survey and medical researchers say back translation.

What is the back translation method in research? One bilingual translator translates a questionnaire, a second translator who hasn't seen the original translates it back, and the team compares the result with the source. Richard Brislin described the method in 1970, and clinical guidelines such as ISPOR's from 2005 include it as one step among several.

What does back translation mean in NLP? In machine learning, back-translation creates training data: a model translates target-language text into the source language, and researchers train a translation model on the resulting pairs. Sennrich, Haddow and Birch introduced the approach for neural machine translation in 2016.