Google introduces neural machine translation to Hindi, translations are now much more accurate

Google is in the process of replacing phrase based machine translation (PBMT) with neural machine translation (NMT) to provide more accurate translations, that scan more naturally for a native language speaker of either language. Neural machine translation was first introduced for Chinese , and then went on to be introduced for Portuguese, Turkish and Japanese. Now, Google has announced that translations to and from Hindi are also handled by neural machine translation.

The old method used phrases to translate, the new method translates the entire sentence at once. The sentences should flow more easily now. The new translations are available wherever Google translate is used, including Android, iOS and the web. A team of bilingual humans helped make the NMT based translations better by continually supplying the system with feedback on the performance of the translations. The translations use the TensorFlow open source framework for machine learning.

Google plans to roll out NMT for all 103 languages available on Google Translate. The translations should be better in translations between Hindi to the other languages, as well as from other languages to Hindi.

The AI powering Google Translate has created a model for understanding languages it has not been trained on before, and can handle translations in new language pairs. Google calls this process  Zero Shot translation .