Beyond Words - Language Blog

Google Translation: Man vs. Machine II

Mere days after my last post about human vs. machine translation, the blogosphere was abuzz with one of Google’s latest moves: AJAX Translation API.

AJAX Translation API is a machine translation tool developed by Google to perform on-the-fly translations for emails, Web sites, and other basic documents on the Internet. Ostensibly, its purpose intersects with one of ALTA’s primary language services-document translation. And in truth, it’s a useful tool for very basic purposes. When quality isn’t a factor, a machine-produced translation can help a user recreate the bare bones or gist of a portion of text. The application also features language detection, a helpful tool that sets Google’s API apart from many other free, Internet-based translation applications.

Still, even Google’s latest effort shouldn’t cause professional translators and translation agencies to lose any sleep. Every now and then we’ll get an inquiry here at ALTA to the effect of “How do I say ‘I love you, Johnny’ in Spanish?” (Not that we mind. And we hope you and Johnny are very happy together!) But the vast majority of the services we provide center around high-quality translation for business documents such as product packaging, contracts, certificates, and employee manuals. And if there was a better way to do it than using intelligent, capable professionals with years of experience, we would utilize it.

I talked with my manager, Rob, to get his personal take on the man v. machine subject.

“If this approach was delivering text our clients could use, we would use it,” he said.
“But we recognize this is not ready for prime time. It will help people get the gist of basic, inbound communication; but, as a form of professional communication, it just isn’t there yet.”

I still give kudos to Google though for the language detection application. Often times our clients come to us, unsure of what language they’re even dealing with. The detection application isn’t perfect, but the software provides a reliability percentage, letting you know how on-the-mark the results may be. I tried out a few examples-it correctly detected Japanese, Arabic and Korean. The application failed to detect the Indonesian, Norwegian and Swedish samples I fed it. And then there are gray areas, like Chinese, which it correctly detected, but didn’t differentiate between Simplified or Traditional Chinese, which are different.

So I know I’m a biased referee. But I don’t care. I’m now upping the score to humans: 2, machines: 0.