The encyclopedia · Software & IT · Technical decision · 2004–2007
Google taught itself to translate by reading millions of human-translated documents.
Google skipped hand-written grammar and trained translation on billions of words of existing translated text, where more data beat better rules.
the move
Machine translation had long been rules based, so every new language required linguists to write grammar and dictionaries and the approach never scaled past a handful of languages.
The answer was not another layer of rules but to stop writing them. Franz Josef Och fed hundreds of millions of words from parallel texts, using United Nations and European Union documents, into a system that learned mappings statistically.
Google's stated reasoning was that the more data fed in, the better the system gets, and its 2007 research showed quality rising with training scale even without new modeling, the lesson that drove the field.
Because parallel text already encodes correct translations, one pipeline could generalize to a new language as soon as enough matching documents existed, without a linguist per language.
why it works
- Parallel corpora are free training data that already encode correct translations.
- Statistical models improve monotonically with data size, so Google's corpus was the moat.
- One pipeline generalizes to any language once enough parallel text exists, instead of needing rules per language.
what transfers
When a problem is too varied to solve by rule, collect historical examples of the desired output and let a model infer the rule, since more data can substitute for more engineering.
what came after
Google Translate went from a handful of languages to dozens and became the default for quick translation, and the data-over-rules principle it demonstrated shaped the field. The statistical approach gave way to neural machine translation around 2016, but the principle that scale and data beat hand-crafted engineering carried straight through.
references
spotted an error? The archive wants to know.