Google has announced a major update of its Google Translate service: the translation will be based on neural networking. Some languages have seen the neural-networking translation since November 2016; others, including Russian, Hindi, and Vietnamese are going to be switched to this method soon. The developers argue that neural networks greatly improve the quality of translation, because thus the machines will be able to analyze not only individual words and phrases, but complete sentences and the context.
“We’re talking about the third generation of machine translation systems. The first was developed in the 1950s-1960s under the Soviet-American spyware confrontation, as everyone wanted to quickly understand what the enemy says. The work was based on the rules developed by linguists, and every sentence had to be translated with different rules applied,” said Barak Turovsky, manager of Google Translate and a USSR migrant to the US, in an interview to Medusa. “The second generation is statistical machine translation. Ten years ago, Google pioneered this technology. A computer uses the same technology that we use in the search engine, searching the Internet for available translations, made by people. Any open source – UN records, news articles – anything that can be found is collected and transformed into a giant database for every language. These data is then analyzed for statistical patterns and thus the engine is ‘studying’ the language. The third generation is neurons; we are talking about technologies that mimic the workings of the brain. Previously, the engine could not efficiently process the phrase more than five words long. The system now looks at the full sentence and tries to find the right options for translation. The result is a much more lively translation.”
Also Barak Turovsky said that work is ongoing to improve the translation with the use of camera and microphone in the “conversation mode” and that plans are made to add the “Tap to Translate” function to all popular online messengers.