Google Translate launches with statistical machine translation

Background

Before 2006, Google did not build its own translation engine. From 2001 it offered translation between eight languages and English using commercial machine translation systems provided by others^[1]. That arrangement covered a small, fixed set of languages and depended on outside technology.

The announcement

On 28 April 2006, Google announced its statistical machine translation approach^[1]. Wikipedia dates the launch of Google Translate as a statistical machine translation service to the same day^[2]. The languages did not all arrive at once: in early 2006 Google rolled out Chinese first and then Arabic^[1].

How the approach worked

Statistical machine translation differs from the rule-based systems that preceded it. A rule-based system encodes grammar and a dictionary by hand; a statistical system is given large collections of text that already exist in two languages and estimates from counts which words and phrases tend to correspond. Wikipedia says the service originally used United Nations and European Parliament documents and transcripts to gather its linguistic data^[2]. Such material suits the purpose: it is translated by professional translators into many languages and published side by side, so the same content exists in matching versions.

Wikipedia also records a structural choice. Rather than translating languages directly, the service usually pivoted through English before translating into the target language^[2]. A request from one non-English language to another would therefore be handled as a translation into English followed by a translation out of English. Pivoting sharply reduces the number of systems that must be built and maintained, at the cost of an extra step in which meaning can drift.

Making it fast enough to ship

The hard problem was speed. Franz Och wrote that the early data-driven system was too slow to be a practical service: translating 1,000 sentences took 40 hours on 1,000 machines^[1]. A year later, the system could translate a sentence in under a second, and with better quality^[1]. That gap — from heavy computation per sentence to a result in well under a second — is what turned a research method into something that could be offered to the public.

Reach

By 2012, Google said Translate had more than 200 million monthly active users and could translate among 64 languages, with more than 92 percent of traffic coming from outside the United States^[1]. The distribution matters as much as the total: most usage was not confined to a handful of Western markets.

Later change: neural machine translation

In November 2016, Google announced a switch to Google Neural Machine Translation (GNMT)^[2]. The statistical systems of the 2006 generation assembled a translation from smaller pieces; GNMT translates whole sentences at a time^[2]. That shift replaced the method the service launched with, while the service itself kept the same role: typing or pasting text in one language and reading it in another.

Why it matters

Google Translate is a useful case because its history is visible in the product. It began as an application of a statistical method on a limited set of languages, was made usable by an engineering effort on speed, scaled by reusing one pivot language, and was later rebuilt around neural networks. Each stage answered a practical constraint rather than a single breakthrough, and the public only saw the result: a free translation box on the web.