Google Neural Machine Translation — The Overnight Jump in Translation

GNMT (Google Neural Machine Translation) is the deep-learning machine translation system Google announced on 27 September 2016. At launch Google switched all of Google Translate's Chinese-to-English translation to GNMT, and on 15 November 2016 (16 November in Japan) it extended the system to eight language pairs to and from English, including Japanese. The November switch is the moment many people in Japan noticed that English–Japanese translation had suddenly started producing readable Japanese.

Until then Google Translate had used phrase-based statistical translation, which split a sentence into phrases, translated them largely independently and picked the statistically most likely combination. Google's Japanese blog recalled that Google Translate had started out as a pioneer of statistical machine translation; how that approach arose is told on the statistical machine translation page. GNMT replaced it with a neural network that treats the whole sentence as one unit.

September 2016: the paper and the Chinese-to-English switch

On 26 September 2016 Google researchers led by Yonghui Wu and Mike Schuster posted "Google's Neural Machine Translation System: Bridging the Gap between Human and Machine Translation" to arXiv. The author list includes Japanese researchers such as Taku Kudo and Hideto Kazawa. The next day, Quoc V. Le and Mike Schuster announced the production deployment on Google's research blog.

According to the paper, simply stacking LSTM layers worked well up to four layers, barely with six and very poorly beyond eight; residual connections were added to get past that limit. Wordpieces strike a balance between the flexibility of characters and the efficiency of whole words, so even a word missing from the vocabulary can be translated as a combination of parts.

The research blog said that, as of 27 September, the Google Translate mobile and web apps were using GNMT for 100% of Chinese-to-English translations, about 18 million a day. It also said GNMT reduced translation errors by more than 55% to 85% on several major language pairs. Both figures come from Google's own evaluation on sentences sampled from Wikipedia and news sites.

November 2016: eight language pairs, Japanese included

In the November announcement Barak Turovsky, Google Translate's product lead, wrote that these eight pairs covered the native languages of around one-third of the world's population and more than 35% of all Google Translate queries, and that "Google Translate is improving more in a single leap than we've seen in the last ten years combined". That is the provider's own assessment. The same day, businesses and developers gained access to the system through an API.

Japan was in the first expansion wave. But many Japanese users noticed the change before the official announcement.

What happened in Japan

According to a long New York Times Magazine article in December 2016, late one Friday night in early November Jun Rekimoto, a professor at the University of Tokyo, noticed posts on social media saying Google Translate had suddenly improved. He tested it himself and wrote up the results on his blog. First he set passages from two published Japanese translations of The Great Gatsby, by Takashi Nozaki and by Haruki Murakami, beside Google's version. Then he translated the opening of Hemingway's "The Snows of Kilimanjaro" into Japanese himself, ran it back into English through Google, and invited readers to guess which version was the machine's. By dawn, the article reports, Google Translate was the number-one trend on Japanese Twitter.

Japanese assessments pointed the same way. In December 2016 Lifehacker Japan wrote that, subjectively, English-to-Japanese output had gone from something like 15 points to 85. Hiromi Nakaiwa, president of the Asia-Pacific Association for Machine Translation (AAMT), wrote in the Japio YEAR BOOK 2017 that rule-based translation built on hand-made dictionaries and rules, and statistical translation, had been mainstream in Japan, and that the November 2016 rollout greatly improved machine translation quality involving Japanese and had a large impact not only on researchers but on society.

How translation work changed

Once neural translation came close to being usable, the question shifted from whether machine translation helps to how to use it. Lifehacker Japan said the new Google Translate was good enough for getting the gist of, say, a business email, but that for professional work a person who can read the original must check the output, since nonsensical translations still turned up. It suggested a workflow for translators in which the machine drafts first and a human then fixes the details.

Google itself stated that machine translation was by no means solved. The research blog wrote that GNMT could still make significant errors a human translator never would, such as dropping words, mistranslating proper names or rare terms, and translating sentences in isolation rather than considering the context of the paragraph or page.

Hope and disappointment in machine translation have a long history; the disappointment of the 1960s is mentioned on the AI winter page. The year after GNMT, Google researchers published the Transformer, which translates with attention alone and no recurrence, and the leading approach changed again. For the wider story see the AI timeline and the Google page.