GB/T 19682-2005Target text quality requirements for translation services (English PDF)

1 Scope

This standard specifies the basic requirements, the special requirements and the other requirements for the quality of a translation, together with the method of assessing and testing that quality.

It applies to translation services, that is, to the relationship between a client commissioning a translation and the provider delivering one.

4 Requirements for the target text

The basic requirements govern what a delivered translation has to be: complete, faithful to the source, correct in its terminology, consistent within itself, and written in a target language that reads as the target language rather than as a transposition of the source.

The special requirements cover what varies by document type. A contract, a patent specification, a set of operating instructions and a marketing text are judged against different criteria, and a single quality definition that ignored the difference would be unusable for all four.

The other requirements cover the material around the text: the treatment of names, of numbers and units, of figures and tables, of headers and page furniture, and of anything in the source that cannot be translated and has to be handled by convention.

5 Classification of errors and assessment

The standard does not attempt to define good translation. It defines error, classifies it, and assesses quality by counting.

Errors are separated by kind - omission, addition, mistranslation, terminological inconsistency, language error - and by severity, with the serious ones being those that change what the text says rather than how it reads.

The assessment then works from a sample rather than from the whole delivery, at a defined rate, and produces a figure that can be compared with a threshold. That is what makes the method usable on a document of a hundred thousand words where nobody is going to check every sentence.

7 Handling of disputes

The final clause covers what happens when the client says the translation is not good enough. It is the clause the whole document is built towards.

It sets out who assesses, on what sample, against which criteria, and what the outcomes are: correction by the provider, a reduction in price, or rejection. What it removes is the argument about whether the translation is any good, and replaces it with an argument about a count of classified errors.

That is a smaller and more tractable argument, and one that a third party can adjudicate without having formed an opinion of the text.

The buyer cannot judge what was bought

Translation is one of the few services where the client is structurally unable to evaluate the deliverable. If they could read the target language well enough to assess it, they would not have needed the translation.

That asymmetry is what makes disputes in this trade so unproductive. The client judges by proxies - fluency, length, how it looks - none of which detect the failure that matters, which is a sentence that reads perfectly and says something the original did not.

So a quality standard here cannot work the way a product standard does. It cannot describe the good outcome; it describes the failures, defines how to look for them in a sample, and specifies what follows from finding them.

It is a procedural answer to a problem that has no substantive one, and it is the right answer: what the parties actually need is not a definition of a good translation but an agreed way to settle whether this one is acceptable.

The buyer cannot judge the product

A client who could evaluate the target text would not have needed the translation in the first place.

That asymmetry is why a national standard exists here at all.

So it defines error, not quality

Good translation cannot be reduced to a rule. What can be reduced to a rule is what counts as a defect and how much each defect weighs.

Errors of meaning - omission, addition, mistranslation, wrong sense of a term - are separated from errors of terminology, of language and style, of formatting and of typography, and weighted differently. A mistranslated clause in a contract and a wrong font in a heading are both defects and are not comparable.

And turns the argument into a procedure

A defined proportion of the delivered text is sampled, the errors found are classified and weighted, and the error density is compared with a threshold.

That is the only way a commercial dispute over a translation is resolved without an expert on each side and no way to choose between them.

Counting errors in a sample, and what that admits

Assessing a translation by drawing a sample and counting classified errors is a method borrowed from manufacturing inspection, and applying it to language carries an admission worth stating.

It admits that no translation of any length is error-free, and that the question is not whether errors exist but how many and how serious. That is uncomfortable for a client who expected the document to be right, and it is the only workable basis for a commercial standard.

It also admits that the assessor is not going to read the whole thing. On a large document nobody does, and a method that pretended otherwise would be honoured by sampling anyway, without the rules that make the sample representative.