Google Translate - Success of Machine Learning

Google announced the new translator algorithm for its Google Translator in late september. And finally we could have to meet it recently. Actually I was not expecting the fast come out of the new algorithm but today we could have a chance to use it and see the results. As far as I could test; it is really successful, even much more successful than before.

Thankfully Google brought the new algorithm for Turkish language translations as well, so I could test it with some complicated English>Turkish, Turkish>English paragraphs and it was looking really professional as if the paragraphs were translated by a professional translator.

Google says; professional translators worked as advisors of the project and they re-coded the algorithm according to the new neural networks machine learning system.

 At present, Google Translate uses Phrase-Based Machine Translation which aims to reduce the language barrier carried by longer sentences and complex contextual insertions in a given language. While this worked to a certain degree of acceptability, it was never a perfect and error-free method of translation.

Simply put, Neural Machine Translation tries to translate whole sentences at a time instead of a word-for-word translation method, which often leads to errors. It uses a broader context to determine what the most relevant translation is. It doesn't stop here, however. It then arranges and adjusts the translation to package it with proper grammar as to sound more human-like.

Best of all, the Neural Machine Translation gets better at translating sentences as people use it more frequently, because end-to-end to learning system is built onto Neural Machine Translation.    [TechTimes]

In order to achieve that much sucess with the Neural Machine Learning you would need an huge database and effective computing of the architecture. No doubt; only Google could do that..

 

Sample translation of a Wikipedia Article

References
Google Translate
Techtimes
Deep Learning Algorithm Wikipedia