Microsoft detailed its results in a blog post this week. Researchers at the company’s labs in the U.S. and Asia said they have achieved human parity when translating the newstest2017 collection of news articles from Chinese to English. The articles are commonly used when testing and benchmarking translation results.
Microsoft hired third-party bilingual human evaluators to assess the suitability of its methodology. The evaluators compared the results of Microsoft’s AI with translations produced by two human linguists. The human translators worked independently of each other to create their renditions of the news stories.
Microsoft achieved the results using deep neural networks. The technique allows it to train AI systems to deliver natural-sounding translations that are accurate and fluent. Deep neural networks enable the AI to consider the broader context in which a translation is occurring. This helps to remove the errors common to conventional statistical translation systems.
The researchers also used dual learning to enhance the AI. This is a way in which the AI can check its work to further verify the accuracy of results. For each sentence translated from Chinese to English, the AI translated the result back to Chinese. This process gives the AI similar checking capabilities to those of human translators. It also provides opportunities for the AI to iteratively refine its output by learning from earlier mistakes.
READ NEXT: Microsoft to provide PCs to teacher who taught IT on a blackboard
Tie-Yan Liu, a principal research manager at Microsoft Research in Beijing, said much of the progress in AI-powered translation “is really inspired by how we humans do things.” The success of dual learning systems is an example of how human techniques can improve automated systems. Another mechanism is the use of deliberation networks, which emulate the way in which people recursively edit and revise their writing.
According to Microsoft, the research lays the groundwork for improvements to the company’s commercial translation services. While the results on the newstest2017 benchmark were based on prototype technologies, they could be refined and released to public cloud consumers. Microsoft said this would create “more accurate and natural-sounding” results on its translation platforms.
There’s no word yet on when the technology will be ready to go live. However, Microsoft confirmed to ZDNet that it intends to bring the system to platforms such as Microsoft Translator “as soon as possible.” It could benefit several of Microsoft’s core services, including Cortana speech recognition, global Office integrations and the company’s just-announced real-time message translation in Teams.
