Machine translation

What machine translation achieves

Machine translation converts texts automatically into another language. Modern systems work with language models and deliver results that are grammatically correct and fluently readable. That is a big step from earlier methods that translated word for word and destroyed the sentence structure doing so.

The limit lies not in grammar but in meaning. A system does not know the product, not the market and not the expectations of the audience. A technical term is translated in its most common meaning, and that is not always the right one for this assortment.

Where it works reliably

• Short factual details: dimensions, materials, care instructions and technical data translate well, because the terms are unambiguous.
• Standard wording: shipping notes, payment information and standard paragraphs repeat and can be checked once and then reused.
• Attributes: colour and size labels are limited and can be checked as a list rather than for every item again.
• A first draft: anyone translating a long text can produce the draft automatically and revise it afterwards. That saves time even though the review remains.

Where the review stays indispensable

The first point is technical terms. A term can sound similar in two languages and mean something different. In retail that mostly concerns legal terms such as withdrawal, warranty and retention of title, where an imprecise rendering creates a false promise.

The second point is statements with effect. A product description claiming a property is an advertising statement and has to be true. A mistake there is not embarrassing but attackable. The third point is mandatory disclosures: withdrawal notice, privacy policy and terms belong reviewed by a specialist in every language, because they carry legal effect.

The fourth point is tone. A brand writing factually in German quickly sounds formal or too promotional in machine translation. The content is right and the text is still not its own.

How the effort stays small

Three measures keep the effort down. First a glossary with fixed terms per language, so the same thing always carries the same name. Second the review by native speakers of the target language, not of the source: anyone who does not speak the language will not recognise an off tone.

Third tiering by risk. Technical data, attributes and standard notes run automatically, marketing texts are revised, and legal texts are reviewed. That three way split prevents the same effort being applied to every text.

A fourth point is labelling. Machine translation is machine generated content, and the labelling question has not been settled conclusively. Anyone labelling visibly is on the safe side. Basics are covered under generative AI.

Frequent questions

• Is machine translation enough? For factual details yes, for anything with effect no.
• What does it cost? Far less than a translation, but the review remains effort.
• How do I spot errors? By terms that exist in both languages but are used differently.