Pivot Language Translation: Why So Much Translation Passes Through English

Source, bridge, target: how pivoting works

A direct translation goes from Polish to Portuguese in one step. A pivoted translation goes Polish to English, then English to Portuguese. Each step is an ordinary translation, done by a person or a system, and the second step only sees the bridge text. It has no access to the original.

Pivoting is old and not specific to machines. Conference interpreting uses relay: when no interpreter covers a rare pair directly, one interpreter renders the speech into a common language and colleagues interpret from that rendering. Publishing has a long history of indirect translation, where a novel reaches a new language through an existing English or French translation. Subtitling for streaming and film often works from an English template file that many language teams translate from, rather than from the original dialogue.

Machine translation adopted the same idea. Older statistical systems, including large public translators, were widely reported to route many language pairs through English. Modern neural systems can be trained to translate between non-English languages directly, and some research models are built for many-to-many translation, but whether a specific product pivots for a specific pair is rarely documented and can change between versions.

Why systems and teams route through English

The first reason is data. Training a translation system takes large amounts of parallel text, meaning the same content in two languages. Pairs with English have plenty: international organizations, software, films and websites translate into and out of English constantly. A pair like Finnish to Korean or Czech to Vietnamese has comparatively little.

The second reason is arithmetic. With 21 languages, there are 420 directed language pairs (21 times 20). Routing everything through one hub language needs only 40 directions: each of the other 20 languages into English and out of English. Building, testing and maintaining 40 systems or 40 translator rosters is far easier than 420.

The third reason is people. It is usually easier to find a freelance translator or reviewer for Turkish to English and English to Swedish than for Turkish to Swedish directly. Even when a direct expert exists, the deadline or budget may not reach them.

What English drops in the middle

English is a convenient hub but a lossy one. It leaves several things unmarked that many other languages must make explicit. Whatever the source expressed, the bridge flattens, and the second step has to reconstruct:

Formal vs informal "you"

Lost in English; the target guesses du or Sie, tú or usted, ты or вы

Singular vs plural "you"

Lost; the target guesses whether one viewer or a group is addressed

Grammatical gender of people

Often lost ("the doctor", "my friend"); the target may default to masculine

Honorifics and speech levels

Japanese and Korean politeness levels collapse into neutral English

Dropped subjects

Languages that omit pronouns get an English "he", "she" or "it" invented, which the target then inherits

Idioms

Either translated literally into English, then literally again, or swapped for an English idiom that is translated word for word

Name spellings

Names get an English romanization first, and the target transliterates the English spelling rather than the original

The mechanics of formality, gender and pronouns are explained in context in machine translation. What pivoting adds is that even a perfect second step cannot recover information the first step threw away.

Errors also compound. If each step gets a sentence wrong some of the time, two steps get it wrong more often than one, and a small mistranslation in the bridge becomes the confident premise of the final sentence. A homograph resolved the wrong way in English, such as "bank" or "right", is translated into the target with no hint that the original meant something else.

Recognizing a translation that went through English

You usually cannot be certain, but some symptoms are telling to a reader who knows the target language:

• English word order or sentence rhythm that sounds translated even where the source language would not produce it.
• English idioms, rendered literally, in a text whose source never used them.
• Formal and informal address switching between sentences, because each was guessed independently.
• Women referred to with masculine forms in a target language that marks gender, where the source made their gender clear.
• Russian, Arabic or Hindi names spelled in the target the way English spells them, rather than by the target language's usual transliteration.
• Units, dates or numbers formatted the American way in a text that went between two metric, day-month-year languages.

None of these proves a pivot. Direct systems make some of the same mistakes, which are catalogued in common machine translation errors. But if several appear together, check the categories in the table above first.

Spot-checking a non-English to non-English translation

The useful trick is to check the places pivoting damages, instead of reading everything with equal attention. You need the source transcript and the translated transcript, ideally with timestamps so you can line them up.

1. Read the source transcript and mark the "flatten points": every formal or informal address, every plural "you", every reference to a person whose gender the source marks, every honorific, and every idiom.
2. Note the names, brands and places, and how each should be spelled in the target script.
3. Open the translated transcript and check each marked point against your list. Is the register consistent? Is the gender preserved? Is the idiom meaningful rather than literal?
4. Check the names one by one against your list.
5. Read three short passages in full for meaning: the opening, the densest part, and the most conversational part.
6. If you find errors at more than a handful of flatten points, plan a full review by someone who reads both languages, rather than fixing only what you found.

The best reviewer reads both languages. If nobody does, two reviewers who each read one of them plus English can work together, accepting that their conversation is itself a pivot. Comparing the translation with an English version can also help a reviewer who does not read the source, though back translation explains why that kind of check misses some errors.

Trade-offs: direct is not automatically better

It is tempting to conclude that pivoting is a flaw and direct translation is always preferable. In practice it depends. For a pair with very little parallel text, a direct system may be weaker than two strong systems chained through English. For human work, a skilled pair of translators working via English can produce better results than one inexperienced direct translator.

What pivoting reliably does is shift where the errors are. A pivoted translation tends to be fluent and semantically close, with damage concentrated in the categories English doesn't mark. That makes it reviewable, if the reviewer knows where to look. The real risk is a team that doesn't know a pivot happened and reviews only for meaning.

There are also limits on what any check can catch. A reviewer working from the transcript sees the words but not on-screen context, such as who is in the shot when "you" is said. Watching the video for the flatten points takes longer but catches more.

Non-English language pairs in mydubly

mydubly works with 21 languages, and the spoken language is detected automatically, so a Spanish video can go straight to Polish, Japanese, Turkish or any other supported language without you choosing a bridge. mydubly's translation step uses a configurable translation engine, and this article does not describe how it handles any particular pair. Treat any non-English pair as one where the checks above are worth doing.

The outputs make those checks practical. A job produces a timestamped transcript in the spoken language and, if you choose, a translated transcript and translated subtitles as SRT and VTT, so a bilingual reviewer can compare source and target directly without going through English. If a reviewer only reads English and the target, run the same file a second time with English as the target and give them both translations. For a 10-minute video, each transcript and subtitle run costs 10 credits (1¢), which is small compared with the review time it saves.

For a dubbed version, the voice speaks the translated text as written, so register and gender errors become audible. Run the flatten-point check on the translated transcript before you publish the dub, because an edited transcript doesn't regenerate the voice; a wrong line has to be replaced in an editor. The video translator page lists all supported languages, and the audio translator works the same way for audio-only files.

Next step

Take one non-English pair you publish in, pull the source and translated transcripts for a short video, and run the flatten-point check above. It usually takes less time than the video's own length and shows quickly whether a full review is needed. For the full translation process around it, see how to translate a video.