Microsoft announced today that it has reached a new milestone in machine translation with a system that can translate sentences from news articles from Chinese to English "with the same quality and accuracy as a person."
According to the researchers, the system reached human parity on the newstest2017 test, a commonly used test of news stories developed by industry and academic partners. To ensure the results were accurate, the researchers also hired outside bilingual evaluators to check the system's translations against references produced by humans.
"Hitting human parity in a machine translation task is a dream that all of us have had," said Xuedong Huang, a Microsoft technical fellow who leads the company's Speech and Language group. "We just didn't realize we'd be able to hit it so soon." Huang was also in charge of the research group that recently reached human parity in conversation speech recognition.
In order to hit the translation milestone, Microsoft says that it employed deep neural networks to train its AI systems, rather than the previous approach, statistical machine translation. This allowed the system to create translation more akin to what humans would naturally produce. The team also took advantage of a mix of training methods, including dual learning, which translates a sentence from Chinese to English, then back from English to Chinese to ensure it makes sense. Deliberation networks, meanwhile, tasked the system with translating the same sentence over and over to edit and revise its own writing.
The results are based on a single test set, which includes around 2,000 sentences from a sample of online newspapers, so it's unclear when or if the system may be ready to use in publicly available translation tools. However, if you've used an online translator with Chinese text, you'll be familiar with the general wonkiness of the results, so any step forward is definitely interesting.
