Microsoft may be getting closer to putting translators out of work. A team of company researchers has developed a machine system that they claim can translate news articles from Chinese to English with the same quality as a human person.
On Wednesday, the researchers called the achievement a major milestone. Many experts had believed designing a computer to match the translation abilities of a human could never be achieved, Microsoft said in a blog post.
“Hitting human parity in a machine translation task is a dream that all of us have had,” said Microsoft technical fellow Xuedong Huang in the post. “We just didn’t realize we’d be able to hit it so soon.”
But are the translations really that good? I’m a Chinese speaker and so I tried the system, which Microsoft has put online. It’s actually impressive and seems to excel at writing translations that sound natural in the English language —something which Google Translate and Microsoft Bing can struggle to do.
However, the system hedges a bit. It’ll give you not one, but two different translations, both of which can vary in accuracy and sometimes include mistakes. (In other words, human proofreaders will still be needed.)
What makes Microsoft’s solution special? It taps into an AI-based technology known as neural networks, or computing systems that have been modeled after biological brains. In recent years, neural networks have attracted growing interest from computer scientists for their capacity to learn like humans can: with practice.
In Microsoft’s case, the company’s researchers trained the AI system to focus on writing fluent and natural-sounding translations by doing the process repeatedly.
“One method they used is dual learning,” Microsoft’s blog post said. “Think of this as a way of fact-checking the system’s work: Every time they sent a sentence through the system to be translated from Chinese to English, the research team also translated it back from English to Chinese.”
This back and forth taught the AI what mistakes it had made and how to fix them. On top of that, the researchers also trained the computing system to translate a sentence of text over and over again until it was polished.
To show the AI system could match the capabilities of a human, Microsoft ran it through a test that judged the results against human-produced translations. This approach involved hiring outside bilingual experts to score the translations from 0 to 100, and then compare the average.
This testing method is maybe a little questionable; it isn’t comparing one individual translation against another, but looking at the overall performance. Nevertheless, the test showed the AI system scoring slightly better —from 68.6 to 69— than the human-produced translations, the highest of which came in at 68.5.
But despite the achievement, the researchers said in Microsoft’s blog post that machine translation is not a solved problem. Their own paper on the technology mentions randomly sampling 500 sentences translated by the AI system and occasionally finding various errors, including incorrect words, bad grammar, and problems with names.
Still, the work represents progress. “What we can predict is that definitely we will do better and better,” said Microsoft principal research manager Tie-Yan Liu in Wednesday’s blog post.
Although the research only covers Chinese and English, the technologies can be useful in other languages or perhaps even beyond translation, Microsoft added.