Translation methodology
In April 2006, Google Translate launched with a statistical machine translation engine.
Google Translate does not apply grammatical rules, since its algorithms are based on statistical or pattern analysis rather than traditional rule-based analysis. The system's original creator, Franz Josef Och, has criticized the effectiveness of rule-based algorithms in favor of statistical approaches. Original versions of Google Translate were based on a method called statistical machine translation, and more specifically, on research by Och who won the DARPA contest for speed machine translation in 2003. Och was the head of Google's machine translation group until leaving to join Human Longevity, Inc. in July 2014.
Google Translate does not directly translate from one language to another (L1 → L2). Instead, it often translates first to English and then to the target language (L1 → EN → L2). However, because English, like all human languages, is ambiguous and depends on context, this can cause translation errors. For example, translating vous from French to Russian gives vous → you → ты OR Bы/вы. If Google were using an unambiguous, artificial language as the intermediary, it would be vous → you → Bы/вы OR tu → thou → ты. Such a suffixing of words disambiguates their different meanings. Hence, publishing in English, using unambiguous words, providing context, or using expressions such as "you all" may or may not make a better one-step translation depending on the target language.
The following languages do not have a direct Google translation to or from English. These languages are translated through the indicated intermediate language (which in most cases is closely related to the desired language but more widely spoken) in addition to through English:
Belarusian (be ↔ ru ↔ en ↔ other);
Catalan (ca ↔ es ↔ en ↔ other);
Galician (gl ↔ pt ↔ en ↔ other);
Haitian Creole (ht ↔ fr ↔ en ↔ other);
Korean (ko ↔ ja ↔ en ↔ other);
Slovak (sk ↔ cs ↔ en ↔ other);
Ukrainian (uk ↔ ru ↔ en ↔ other);
Urdu (ur ↔ hi ↔ en ↔ other).
According to Och, a solid base for developing a usable statistical machine translation system for a new pair of languages from scratch would consist of a bilingual text corpus (or parallel collection) of more than 150–200 million words, and two monolingual corpora each of more than a billion words. Statistical models from these data are then used to translate between those languages. To acquire this huge amount of linguistic data, Google used United Nations and European Parliament documents and transcripts. The UN typically publishes documents in all six official UN languages, which has produced a very large 6-language corpus. Google representatives have been involved with domestic conferences in Japan where it has solicited bilingual data from researchers.
When Google Translate generates a translation proposal, it looks for patterns in hundreds of millions of documents to help decide on the best translation. By detecting patterns in documents that have already been translated by human translators, Google Translate makes informed guesses (AI) as to what an appropriate translation should be.
Before October 2007, for languages other than Arabic, Chinese and Russian, Google Translate was based on SYSTRAN, a software engine which is still used by several other online translation services such as Babel Fish (now defunct). From October 2007, Google Translate used proprietary, in-house technology based on statistical machine translation instead, before transitioning to neural machine translation.
