Microsoft is developing the company’s deep learning neural networks for improving the translation of different languages by tapping further into the capabilities of AI technology. Such developments could potentially increase the accessibility of foreign information and news, allowing such material to be easily shared around the world.
The technology giant has recently announced its researchers in the US and Asia have successfully developed an AI translator which is able to perform translations with the same level of competency as a professional human translator. This new AI tool, which has been tested using a well-accepted, open-source method using 2017 Chinese news stories, is said to have “achieved human parity”. Similar testing methods are also used by other tech companies for benchmarking purposes of similar research activities.
The technology giant is now at the stage of graduating its increasingly robust AI technology into the next stage of curriculum. According to assistant managing director of Microsoft Research Asia, Ming Zhou, following the success of the AI translating tool on old news stories, the research team plans to extend the success to a wider set of content and real-time stories. After reaching this milestone, the company will then apply the AI tool into Microsoft Translator.
Zhou also explained that unlike other AI tasks, which generally involve pure pattern recognition, machine translation is by far more complicated because words are more of an art rather than a science. Different people may choose different methods of expressing something, and it is difficult to determine which way is the best.
In order to help its AI learn and improve from mistakes, the company actually applies the same method as if teaching a human being. Similar to how a human learns, the AI at Microsoft receives feedback and then corrects its mistakes during their next attempts. This is a major factor of the success of the AI translation tools developed by Microsoft.
For example, Microsoft applies the “dual learning” method - the AI is told to perform a translation from Chinese to English, followed by an immediate translation of the text back to Chinese, to test how much of the original meaning has been preserved in the process. A different teaching method, “deliberation networks”, asks the AI to perform repeated translations of the same text and provides it with human feedback during each iteration.
Another strategy is called “agreement regularisation” which gets the AI to perform backward and forward reading of a sentence, following by translating the sentence in both orders. Depending on whether the translation results match, the capability of the AI to convey the correct information is put to the test. Very often, translation tools have a hard time in properly arranging a sentence, which results in a jumbled mess of words.
Although the Microsoft AI translation tool is still in relatively early stages, the company is making good progress in using AI technology to improve communication of people around the globe by removing the barrier of language.
Contact WM Reply today for some friendly advice on how Microsoft’s new services and tools could be applied to significantly improve the productivity of your business.