Google translation utilizes artificial intelligence aiming at human-level translation accuracy

You can use it on smartphone, PC or offline,Supports over 100 languagesGoogle's translation service "Google translation"Announced the" GNMT "system utilizing artificial intelligence (AI) that enables more natural translation at the timing that celebrates its 10th anniversary.

Research Blog: A Neural Network for Machine Translation, at Production Scale
https://research.googleblog.com/2016/09/a-neural-network-for-machine.html


Google's AI translation system is approaching human-level accuracy - The Verge
http://www.theverge.com/2016/9/27/13078138/google-translate-ai-machine-learning-gnmt


With AI research with several major IT companiesI just got a partnershipGoogle announced the AI-based translation system "GNMT (Google Neural Machine Translation)". In Google Translate, a system called PBMT that machine-translates on a phrase basis has been adopted so far, but it is not a method of mechanically translating sentences for each word or phrase, but capturing the entire sentence as one translation unit It is said that it is GNMT that can do.

According to Google researchers, "The advantage of GNMT's" capturing the entire sentence as a translation unit "approach is that fewer engineering design choices are required than PBMT." In the first translation that actually used GNMT, it seems that translation accuracy as good as existing Google translation was seen. Furthermore, by repeatedly translating, GNMT seems to be able to achieve both excellent translation and quick translation speed. According to Google, Google translation will be able to reduce translation mistakes by as much as 55 to 85% by using GNMT.

According to Google, GNMT has reached a level where human-level translation is possible in some cases. The following graphs compare the three translations by humans, GNMT, PBMT in six grades and compare them. Although the translation accuracy is highest for human beings, the translation is mostly translated by humans, but when translating from French to English or from English to Spanish It seems that there is not such a big difference between humans and GNMT.


The following GIF image visualizes the process of translating Chinese sentences "knowledgeable ability (Knowledge is power)" into English by GNMT. First of all, GNMT 's network encodes Chinese words into a list of vectors (one - dimensional array). Individual vectors show that "words have been read so far". When all sentences to be translated are read, the decoder will begin to generate English sentences word by word. The blue line extending to connect the generated English word with the original Chinese word indicates that "how much attention the decoder took when generating translated words".


The following table compares the results of actually translating the leftmost Chinese sentences with PBMT, GNMT, and three human patterns. Compared with PBMT, it seems that GNMT is approaching more natural sentences.


In order to increase the accuracy of GNMT, Google has already implemented GNMT in Google Translate with limited translation from Chinese to English. There are about 18 million translations from Chinese to English per day when combined with web version and mobile version, and it seems that all these are handled by GNMT. Also, within the next few months GNMT will be used when translating in other languages.

For further information on Google's research on GNMT(PDF)paperIt is written in.