Recently, Google Translate has upgraded its kernel. According to the official Google data, Google Translate's Google Neural Machine Translation (GNMT) system uses the most advanced training technology available, thus improving the level of machine translation and reducing translation errors by another 55%-85%.
More than a decade ago, Google released Google Translate, an early phrase-based statistical machine translation that breaks down input sentences into words and phrases and then translates them independently.
The disadvantages of this type of translation are obvious: the otherwise complete information in the sentence is fragmented and cannot be expressed coherently. This phenomenon is particularly evident in the case of English-Chinese translation.
And Google Neural Machine Translation translates the input sentences as a whole.
Taking the Chinese and English translation as an example, Google Neural Machine Translation first encodes this Chinese word into a list of vectors, where each vector represents the meaning of all the words that have been read so far (encoder "Encoder").
Read the whole sentence and the decoder starts working - one word at a time for the English sentence (decoder).
In order to generate the correct translated word at each step, the decoder focused on the weight distribution of the Chinese vectors most relevant to the encoded English words generated.
At the time it was first proposed, neural machine translation systems were at the same level as phrase-based translation systems on medium-sized data sets.
Now, Google, for its part, says that by allowing neural machine translation to overcome many of the challenges of working on very large data sets, it has built a system that translates better in terms of both speed and accuracy.
Currently, the Google neural machine translation system has been put into Chinese-English translation.
Now, the mobile and web versions of Google Translate's Chinese and English translations are fully translated using neural machines - about 18 million translations per day.