Google Translate is being improved by artificial intelligence, the IT company reported on Tuesday. The innovative software, called Google Neural Machine Translation is almost as good as a human translator.

As of yesterday, Google Translate has been improved through a new translation system, based on Recurrent Neural Networks (RNNs), which uses state-of-the-art training techniques in order to obtain the best quality translation that a program has ever done up to the present moment. Google Neural Machine Translation (GNMT) is currently available for translations between English and Chinese, of which the IT giant says it processes about 18 million every day.

Google has been using RNN for a few years now. An RNN is a type of artificial neural network that can use the information already achieved and memorized to perform arbitrary sequences of inputs. Before using RNN, the translator software used by Google was based on Phrase-Based Machine Translation, which breaks a sentence into words and expressions that are mostly independently.

After the testing and improvements in the past years, Google has managed to refine NMT to an accuracy almost as good as a translation made by a human being.

To notice the difference, the Research Google Blog gives an example:

 

sep-9

 

The difference is remarkable. Still, the NMT system is far by being perfect and, according to the researchers, it can still make errors that a human would never be able to make. The vulnerabilities mentioned in the report include dropping some words, mistranslating some other words and not translating phrases independently rather than taking the context into account. “There is still a lot of work we can do to serve our users better”, the blog post says.