Last year Microsoft announced the release of its Neural Network based translation system for 10 languages: Arabic, Chinese, English, French, German, Italian, Japanese, Portuguese, Russian, and Spanish. Today, Korean is being added to the list.
Neural Network translation uses the full context of a sentence to translate words based not only on a few words before and after it, but on the full sentence, generating more fluent and more human sounding translations. This new AI-powered technology delivers the most significant improvement in machine translation quality since statistical machine translation became the industry standard 10 years ago.
Thanks to these improvements in quality and fluency, translations are the closest they have ever been to human generated ones.
HOW IT WORKS
At a high level, Neural Network translation works in two stages:
- The first stage models the word that needs to be translated based on the context of this word (and its possible translations) within the full sentence, whether the sentence is 5 words or 20 words long.
- The second stage then translates this word model (not the word itself but the model the neural network has built), within the context of the sentence, into the other language.
Neural Network translation uses models of word translations based on what it knows from both languages about a word and the sentence context to find the most appropriate word as well as the most suitable position for this translated word in the sentence.
One way to think about neural network-based translation is to think of a fluent English and French speaker that would read the word “dog” in a sentence: “The dog is happy”. This would create in his or her brain the image of a dog. This image would be associated with “le chien” in French. The Neural Network would intrinsically know that the word “chien” is masculine in French (“le” not “la”). But, if the sentence were to be “the dog just gave birth to six puppies”, it would picture the same dog with puppies nursing and then automatically use “la chienne” (female form of “le chien”) when translating the sentence.
Here's an example of the benefits of this new technology used in the following sentence: (one of the randomly proposed on our try and compare site: http://translate.ai)
M277dw에 종이 문서를 올려놓고, 스마트폰으로 스캔 명령을 내린 뒤 해당 파일을 스마트폰에 즉시 저장할 수 있다.
Traditional Statistical Machine Translation would offer this translation:
“M277dw, point to the document, the paper off the file scan command Smartphone smartphones can store immediately.”
Neural Network translation, in comparison, generates this clear and fluent sentence:
“You can place a paper document on M277DW, and then save the file to your smartphone immediately after the scan command.”
The Neural Network translation systems are available for you to use through many entry points:
- Browser: We’d love your feedback on the new Neural Network Korean translation system vs. the legacy statistical one! Visit our try & compare site: http://translate.ai
- Microsoft Translator live feature: When using our new personal universal communicator feature, Microsoft Translator live, neural translations will also be used. For instance, if you use this feature to translate a live presentation from any of the nine supported speech languages to any of the 11 NN-powered translation systems, subtitles will be delivered using Neural Network technology: http://translate.it
- Instant Messages in Skype: Translate instant messages (from and to any of these 11 languages) using the Skype Translator feature in Skype desktop and Skype Preview for Windows 10.
In addition, developers can easily integrate Microsoft Translator Neural Network systems by using the category “generalnn” in their API calls. There is no extra cost in using our neural network models vs. the existing statistical ones so don’t hesitate to use them!
For speech translation projects, the Microsoft Translator speech API already uses neural network translations from any of our 9 speech translation languages to all the 11 neural network powered languages.