Machine translation

Machine translation, usually shortened to MT, is software that translates text from one human language into another with no person doing the translating. A translator may still edit the result, which is called post-editing, but the first draft comes from the machine.

Three generations, then language models#

Machine translation is older than the personal computer. Wikipedia's history of the field records a public demonstration of the Georgetown-IBM experiment in 1954, by a Georgetown University team led by Michael Zarechnak.

Early systems were rule-based. People wrote dictionaries and grammar rules for each language pair, and the software applied them. From the late 1980s, as computing became cheaper, research turned to statistical models, which learn which phrases tend to translate into which from large collections of texts and their translations.

Both have since been superseded by neural machine translation and by large language models, in Wikipedia's words. A neural system learns the whole mapping from examples instead of from rules or phrase tables, and is covered in its own entry, neural machine translation. A general-purpose language model can also be asked to translate in a prompt; Wikipedia calls the approach promising but more resource-intensive than models built only for translation.

What machine translation needs from you#

Every generation shares one weakness: the system only knows what it is given. In a worked example, a booking site's button says Book. Given that single word, an engine has to guess whether it is a verb or a noun, and in German the two are different words: Buch is a book you read, and a reservation button needs Buchen. A product name that is also an ordinary word has the same problem, and so does the choice between formal and informal address, which English does not mark at all.

So the quality of machine translation on a website depends on what goes in with the text, such as the terms that must not vary and the form of address the brand uses, and on a person reading the result on the page before visitors do. Is Google Translate accurate? looks at the evidence on accuracy, and machine vs human translation at when each is the better choice.

Machine translation in LangRelay#

LangRelay's machine translation uses large language models. Every request carries the language variety the Site asked for, such as Dutch (Netherlands), the Formality register you set for that language, and, on paid plans, the glossary terms that occur in the text, including terms to keep as in the source. It starts on its own when a page is first read, when an edit arrives and when you add a language, and it reuses any translation that already exists for the same text instead of producing a new variation. By default a translation goes live as soon as it is ready, and free checks hold a page with untranslated text, the wrong language or a changed price. To approve every page yourself, choose Ask me first.