Machine Translation of Chinese to English from Microsoft
For me, one of the more enjoyable aspects of traveling in Belarus is that of the language. While a high percentage of the people there speak the national language of Belarussian, even more speak Russian. I am not fluent in Russian, but can obtain food and lodging, and transport and directions to sites of interest. Chinese? That’s a different story!
Researchers at Microsoft have created the first machine translation system that translates at a human level (human parity) from Chinese to English. As you can imagine this was no easy task. In fact over many years of effort the task proved to be so difficult that three separate teams worked to make the system more fluent and accurate, achieving a 5.1% error level. That’s accurate 94.9% of the time!
It is interesting to note that the various methods used to improve the system – dual learning, deliberation networks, joint training, and agreement regularization – enable functional learning similar to how humans would approach the same task.
“Much of our research is really inspired by how we humans do things,” said Tie-Yan Liu, a principal research manager with Microsoft Research Asia in Beijing, who leads a machine learning team that worked on this project.
For example, dual learning is a way of fact-checking the system’s work. Every time a sentence is sent for translation from Chinese to English, the researchers would also translate back from English to Chinese. This is similar to what a person would do to make sure that the translation received is accurate. When done within the translation system it enables the system to learn from its own mistakes. Dual learning can be used to improve results in yet other AI tasks, as well as language learning.
Xuedong Huang, a technical fellow in charge of Microsoft’s speech, natural language and machine translation efforts, said that achieving human parity in translation between the two languages, one of the most challenging natural language processing tasks, was a major milestone.
“Hitting human parity in a machine translation task is a dream that all of us have had,” Huang said. “We just didn’t realize we’d be able to hit it so soon.”
The research and work required for the technical aspect is compelling. However, this achievement was especially gratifying because of the possibilities it has for helping people understand each other better.
“The pursuit of removing language barriers to help people communicate better is fantastic,” he said. “It’s very, very rewarding.”