While machine to understand language at this scale is a bit more complicated than speech recognition, as advances in artificial intelligence (AI) and speech recognition have enabled digital assistants to find their ways in the homes where they help consumers with everyday tasks, like controlling the smart home devices.
On the other hand, machine translation of a web page or news article renders it more of a difficult scribbles, which gives rather general view on the story, but always impossible to make any deeper sense.
To solve the puzzle, Microsoft tested the machine translation system repeatedly on a sample of about 2,000 sentences from various online newspapers, with the results compared to actual individual’s translation and even went as far as hiring bilingual language consultants to verify the accuracy.
Surprisingly, the researchers were able to achieve this milestone, despite the fact that machine translation of text input remains a difficult area for decades now.
Albeit, the researchers added their own trained models to the system to improve its accuracy – which equate to how individuals could go over their work time and again to improve on it.
The methods they employed include: dual learning for fact-checking translations; agreement regularization, which is capable of generating translations by reading sentences both left-to-right and right-to-left; and deliberation networks, to repeat translations and refine them; with new techniques like joint training, which will help to boost English-to-Chinese and Chinese-to-English translation systems.
The researchers, however expressed reservations about the system and cautions that the technology isn't yet ripe to be commercialized into Microsoft’s products.
If you wish to test run the new translation system, you can access it on Microsoft’s translate website here: https://translator.microsoft.com/neural.
No comments