Microsoft have created a translation algorithm that translates Chinese sentences to English as well as human translators do.
Translating Chinese sentences into English has been a difficult problem in the past. Some languages align well with the English language, making them easier to translate. As Chinese has a different sentence structure, translating from this language is a challenge for many classical translation algorithms. Thanks to neural machine translation, a technique that has created excellent results in the last couple of years, Microsoft have manged to get their machine translated sentences on par with human translated sentences.
Microsoft's researchers used two existing approaches for neural machine translation: dual learning, and deliberation networks. With dual learning, every translation the system created from Chinese to English was translated back to Chinese to verify that the outcome would mean the same. Humans perform the same thing when manually translating to verify that their translation is correct, and machines seem to benefit from using the same technique. Deliberation networks are a way of revising generated sentences. This technique is again inspired by humans: when we translate a sentence we don't immediately write out the full translation but instead revise our initial translation to come up with a better sentence.
Microsoft have also developed two new techniques to improve their translation algorithm. Their joint training technique augments the training set by translating English sentences to Chinese, which are then translated back to English. Both translation systems are improving using this technique. They also created agreement regularization: they let two algorithms read the Chinese sentence from right to left, and from left to right. If both algorithms came up with the same translation this was considered to be a strong indication that the translation was correct.
It is important to note that Microsoft reached this 'parity milestone' by letting bilingual language consultants compare their created translations to the 'golden translation'. Language translation evaluations compare the output a machine generates to a single sentence translated by a human translator. A downside of this approach is that there are no perfect algorithms for comparing the intrinsic meaning of the translated sentences to see that they mean the same. This is a problem that all translation researchers face, including Google who picked an algorithm that performed weaker on the algorithmic evaluation, but stronger on human comparison as evaluation.
The last few years the machine translation community obtained massive gains in translation algorithms by harnessing the power of neural networks. Google, Facebook, and Microsoft are all in the race to develop the best translation algorithm. Microsoft offers a comparison of classical and neural translation algorithms on this website, although their system that performs with parity to human is not listed here.
