Translation quality control: what can be counted and what has to be read

Translation quality control means separating two questions that are always asked together: does this text say what the source said, and is this text a good text in the target language. Each has its own instrument, and asking them at the same time is exactly how errors get through.

  • Lesson 6 of 6
  • Intermediate
  • Free, no signup
Translation

Checks that find something, and checks that only feel like checking

Quality control sheetbefore delivery

Finds something

  • A bilingual pass on numbers, names and negation only
  • Counting the different forms of one term across the document
  • Reading the target text without the source
  • Back-translation, on the high-risk sentences only
  • A second reader, for a text that gets printed or signed

Only feels like checking

  • Reading your own work once more, straight away
  • A spell checker as the only gate
  • Asking the same model that produced the translation
  • A quality score a model gives its own work
  • Back-translating the whole document instead of every other check

The second column is not a list of useless work; it is work that gets treated as a quality gate on its own and is not one. A spell checker is necessary and not sufficient.

Last checked: Facts and tool names in this lesson are re-checked against their sources on this date.

What translation quality means, and who decides it

Translation quality is not a property of the text, it is the relation between the text and the job it has to do. An internal guide that only has to be understood and a contract clause that is going to be signed hold two different standards, and measuring both with one ruler is either a waste of time or a risk taken.

Underneath that relation, though, error typologies genuinely exist and have been built. The MQM typology, whose core is public and readable, separates errors into several dimensions: terminology, accuracy, linguistic conventions, style, locale conventions, audience appropriateness, and design and markup. The value of that list for a working freelancer is not the naming, it is one practical split: some of these errors can only be seen with the source beside you, and the others can only be seen without it.

A dropped sentence, a negation turned over, a number changed, something translated that should not have been: these cannot be seen except by comparison. Grammar, punctuation, register, and the sentence that is correct but that no native speaker would write that way: those are seen by somebody who has not seen the source. A reader with both texts in front of them forgives a clumsy target sentence, because they know what it was meant to say.

One word about standards, since this is where the misunderstandings gather. ISO 17100, titled "Translation Services: Requirements for translation services", sets requirements for the processes and resources needed to deliver a quality translation service, and its scope explicitly puts the use of raw machine translation output plus post-editing outside itself; that work has a standard of its own, covered in the translation workflow lesson. So the first quality decision, before any review at all, is knowing which of the two your job falls under.

Two mirrored glass documents with a red caliper across them, a green reading lamp and a blue empty checklist

Two passes, each of which sees half the errors

The difference is not in strictness, it is in what you put in front of them.

The monolingual pass

  • The target text only, without the source
  • Fluency, register, consistency
  • In a fresh conversation, preferably a different model
  • Is not allowed to say whether the meaning is right

The bilingual pass

  • Source and target side by side
  • Omission, addition, negation, numbers, names
  • Has no permission to comment on style
  • Output: a candidate list with evidence from the source

The order matters too: the bilingual pass comes first, because repairing an omission can change the sentence, and the fluency pass has to run over the final text.

What back-translation actually catches

Back-translation means giving the target text to somebody who has not seen the source and asking them to render it back into the source language, then setting the two source texts side by side. Where it is taken seriously is clinical research and standardised questionnaires, and the reason is plain: there, one word out of place corrupts the data.

What it catches is specific and not small: a flipped negation, a wrong number, a dropped sentence, a concept that shifted slightly in the target. These are exactly the class that cannot be seen without comparison, and back-translation surfaces them, because the returning text is no longer saying what the source said.

What it does not catch is stated less often. Back-translation says nothing about style: a dry, lifeless target text can come back looking exactly like the source. Worse, there is one situation in which it manufactures an error of its own. The back-translator is also a person, and when a target sentence is clumsy or ambiguous they repair it without noticing and return the repaired version. The result is a healthy looking text that has hidden the real problem. This is why back-translation is an instrument of comparison and not a certificate of health.

Our position: run back-translation over the high-risk sentences rather than the whole document. A dosage, an amount, a date, a contractual undertaking, a safety warning. Ten sentences that have consequences if they are wrong are worth a back-translation; three hundred sentences that only have to be understood are not.

Consistency: the one thing you can actually count

Most of what matters in translation quality takes judgement and cannot be counted. Consistency is the exception: it can be counted, and because it can be counted it should be, rather than guessed at. One concept carrying one name on page three and another on page nineteen is an error no reader ever reports and every reader feels.

We ran exactly that over our own text, and the result taught us something more useful than the number. We wrote a check that compares the sentence count of each Persian passage against its English counterpart; a difference in sentence count usually means the writer merged two sentences or split one, and that is the place to look. Across the 2,827 long passages of this section, 91.7 percent matched and 93 passages differed by two sentences or more, the largest gap being 28 against 15.

That 28 was too large to believe, and it was not true. Our check counted sentence-final marks, and to avoid mistaking a decimal point for a full stop it inspected the digits either side of the dot, but it only recognised Latin digits as digits. In Persian text, where numbers are written with Persian digits, every decimal manufactured a phantom sentence. Once Persian and Arabic-Indic digits were added to the definition of a digit, the same measurement came out at 95.5 percent matching, and the passages differing by two or more fell from 93 to 9. So 84 of the 93 findings were the tool's own error.

The practical lesson: before you trust what a check reports, test the check itself once, and look at the largest gaps in the list by eye, because a tool's error usually gives itself away there rather than in the middle of the distribution. And the main boundary: even a correct check produces a list of candidates, not a list of errors. The 128 passages that survived the corrected check are mostly a writer's right decision to turn one long Persian sentence into two English ones. Term-level consistency follows the same logic and was measured by counting over this same corpus in the specialized translation lesson.

One mechanical check over 152 files, before and after fixing itself

  • 2,827 passages compared
  • 93 differed by two sentences or more, with the faulty counter
  • 9 the same gap, after the fix

The same data, the same day, two results. The only difference was that the counter did not treat a Persian digit as a digit and read a decimal point as the end of a sentence.

The second reader rule, and when there is no second reader

You cannot revise your own translation, and the reason is not carelessness. As you read, you read your intention rather than what is on the page; the sentence that was complete in your head stayed incomplete in the text, and your eye substitutes the mental version. More care does not remove this error, because the error is a product of familiarity.

So the reviser has to be another person, and that person needs three things they are usually not given: the source text, the brief (who this text is for and what it has to achieve), and the glossary with the list of things that are not translated. Revision without those three turns into two people's taste arguing over sentences.

And one thing has to be said to the reviser explicitly: they are allowed to say this is good, leave it alone. A reviser who feels they have to find something will find something, and what they find is usually their own personal style. Changing without cause has a price too: every hand laid on a text that was already right is a fresh chance to break it.

If you work alone and there is no second reader, you have two real substitutes. First, distance: put the text down and read it tomorrow; familiarity fades overnight, and it is the only route we know for making the eye see the text again. Second, changing the medium: print it, or read it in a different typeface and size, or read it aloud. The sentence you stumble on out loud is the sentence your reader stumbles on too. There is a third technique that works for spelling and numbers: read the text from the end to the beginning, passage by passage, because the narrative is no longer carrying your eye along.

When it is worth paying twice for one text

A second review costs money and is not always worth it. The arithmetic is also not the one usually done. The cost of a bad translation is not the cost of fixing it; it is the cost of what that text did up to the moment somebody saw the error. A printed brochure cannot be edited, a signed contract cannot be taken back, and a product page that sold nothing for three months does not get those three months back.

So the criterion is not how important the text is, it is how reversible it is. There are four places where we always recommend a second review: anything that goes to print, anything that gets signed, anything published in public with somebody's name under it, and any text on the basis of which a person does something with their body or their money.

And the places where one pass is enough, and insisting on more is waste: internal correspondence, a text that only has to be understood, a draft that is going to change again anyway, and a translation made in order to decide something about the text itself. The most honest thing you can tell a client is that this text does not need a second review, and that one sentence builds more trust than any discount.

The fast path, with AI

What most people do with a language model is hand it the source and the translation together and ask how good it is. The answer is always a polite paragraph with a few stylistic remarks and no real errors in it. The reason is simple: a model with the source in front of it forgives a clumsy target sentence because it understands what was meant, exactly as you do. The fix is two separate passes, each deliberately given only half of the information.

  1. Break the text into numbered segments. Give each paragraph or each sentence a number; without numbers the model's output turns into a narrative that you cannot check row by row.
  2. Run the first pass with both texts, but close its scope tightly: accuracy only. Every time you let it talk about style as well, the list fills with taste and the real error gets lost underneath.
  3. Run the second pass in a fresh conversation and do not enter the source at all. Use a different model if you can, because a model that translated or edited the text sees its own work as smoother than it is.
  4. Put the two lists side by side and decide yourself. Look first at the rows that appear in both lists; that is usually where a genuine ambiguity in the source sits and where the translator slipped on it.

Copy-ready recipe

PASS 1, bilingual. Give it both texts.

Role: accuracy reviewer. The numbered source text and target text follow below.
Report these categories only: omission, addition, a negation turned over, a different number or date or amount, a different proper name, and anything on the do-not-translate list that has been translated.
Say nothing about style, fluency, tone or word choice. Even if a sentence is bad, do not report it if its meaning is right.
Output, a table only: number | category | source phrase | target phrase | why
If you find nothing in a segment, leave it out of the table.

PASS 2, monolingual. Fresh conversation. Give it the target text only.

Role: native reader of {target language}. You are seeing this text for the first time and there is no other text.
Report these categories only: a sentence a native speaker would not phrase this way, a register out of step with the rest of the text, a term that appears in two different forms in two places, and a sentence that has to be read twice to be understood.
Make no guess about whether this is a translation, and give no opinion on whether the meaning is correct.
Output, a table only: number | category | phrase | suggested form

Before you trust the output: Neither pass proves correctness; both produce candidate lists and the decision belongs to the reviser. The second pass goes wrong in exactly the case where the text is a good translation whose tone does not match the model's expectation, so do not accept every suggestion it makes. And one practical constraint: before pasting a client's text into any service, weigh the confidentiality the same way as everywhere else on this path; sending a text to be checked is the same disclosure as sending it to be translated.

AI in this kind of work

In quality control a language model does one thing better than anyone: it does not get tired. Comparing two hundred rows in a row is exactly the job a person loses somewhere around row one hundred. But one rule holds everything up: whoever did the work is not the reviewer of that work. The rule applies to a human translator and to a model alike, and every tool below is introduced on the assumption of that separation.

Tools that actually help

  • Gemini For the bilingual pass over a long document, because its Flash class is cheap and this pass is mechanical rather than a matter of judgement. Google's own page says the app is available in over 230 countries and territories, and Iran is not on that list.
  • Claude For the monolingual pass, which is judgement rather than comparison: which sentence a native speaker would not phrase this way. Its own privacy page says chats are not used for model training by default, which matters when the text you are sending for review is a client's.
  • Mistral Large 3 For when the text must not leave the building at all. Its licence is Apache 2.0 with open weights and full commercial use, which means you can run it on your own hardware and the confidentiality question disappears. In exchange, Mistral publishes no complete language list for this model.

Where it backfires

The risk specific to this topic is the score. Ask a model how good this translation is out of a hundred and you get a number; that number has no scale behind it, asking twice gives two answers, and worst of all it looks like a measurement once it lands in a report to a client. Typologies like MQM exist precisely so that quality decomposes into demonstrable error categories instead of one figure. Our position is simple: ask the model for a list, not a score, and never write a number into a report when you cannot say where it came from. The second risk is a model reviewing its own work: inside the same conversation it sees what it wrote as sounder than it is, and that is what makes the second pass in a fresh conversation necessary. The third is the standing confidentiality question; the policies are not the same. Google's privacy page for the Gemini apps says human reviewers see some of the data and reviewed chats are retained for up to three years, while Claude's data page says chats are not used for training by default.

Sources: MQM Core typology: error dimensions and definitions Google: Gemini Apps Privacy Hub (human reviewers, retention) Google: where Gemini Apps are available Anthropic: is my data used for model training

Where this advice stops

The measurement in this lesson is over our own text and its boundaries have to be stated plainly: these texts are written separately in four languages rather than translated, so what was counted is an author's decision and not a translator's behaviour, and on a real translation obliged to follow the structure of its source the differences take another shape. No error rate is declared here and none can be, because our check finds candidates rather than errors. We have claimed nothing about the text of ISO 17100 itself beyond the title and scope printed in the catalogue entry; we have not read the standard. Two things are deliberately outside this lesson: monolingual editing, which is its own craft and lives in the editing and proofreading lesson, and certified or sworn translation, which is a legal permit rather than a quality system; no amount of review substitutes for it.

From our own work

We wrote a consistency check and the first thing it found was a fault in itself. It did something simple: compare the sentence count of each Persian passage against its English counterpart, over 2,827 long passages from the 152 lesson files of this section. The first report said 91.7 percent matched and 93 passages differed by two sentences or more. The largest row in the list was 28 Persian sentences against 15 English ones, and that single row is what made us doubt the whole table. The cause was that, to avoid mistaking a decimal point for a full stop, the counter looked at the digits around the dot but recognised only Latin digits as digits; in Persian text every decimal number manufactured a phantom sentence, and the passages carrying more numbers showed the largest gaps. Once Persian and Arabic-Indic digits were added to the definition, the same measurement came out at 95.5 percent and the rows differing by two or more fell from 93 to 9. Arabic came out at 97.6 percent and Turkish at 94.4. So 84 of the first 93 findings were not in the text at all. This is why we test any checking tool against a deliberately broken input before trusting it, and it is what separates a quality report from a printable number.

Real follow-up questions

Is back-translation needed on every project?

No, and running it as a routine eats more budget than it finds errors. Its place is text where one wrong word has consequences: a dosage, an amount, a contractual undertaking, a safety warning. For everything else a bilingual pass over numbers, names and negation does the same job more cheaply.

What do we do if there is no second reader?

Buy distance and change the medium: read it tomorrow, on paper or in a different typeface, and once out loud. These reduce your familiarity with the text, which is the thing that opens the eye again. For numbers and spelling, reading from the end to the beginning works, because the narrative is no longer pulling your eye forward.

Can you get a translation quality score out of AI?

You get a number, but that number is not a measurement: there is no scale behind it and asking twice gives two answers. Ask for a list instead, categorised and with evidence, so it can be accepted or rejected row by row. A number you cannot trace has no place in a client report.