Google Announces Improvements To Translation System Google says that with certain languages, its new system — dubbed Google Neural Machine Translation — reduces errors by 60 percent. For now, it only translates from Mandarin Chinese into English. But the company plans to roll it out for the more than 10,000 language pairs now handled by Google Translate.
NPR logo

Google Announces Improvements To Translation System

  • Download
  • <iframe src="https://www.npr.org/player/embed/496442106/496442107" width="100%" height="290" frameborder="0" scrolling="no" title="NPR embedded audio player">
  • Transcript
Google Announces Improvements To Translation System

Google Announces Improvements To Translation System

Google Announces Improvements To Translation System

  • Download
  • <iframe src="https://www.npr.org/player/embed/496442106/496442107" width="100%" height="290" frameborder="0" scrolling="no" title="NPR embedded audio player">
  • Transcript

Google says that with certain languages, its new system — dubbed Google Neural Machine Translation — reduces errors by 60 percent. For now, it only translates from Mandarin Chinese into English. But the company plans to roll it out for the more than 10,000 language pairs now handled by Google Translate.

KELLY MCEVERS, HOST:

And now for something back on the ground. Web-based translators have made it much easier to understand things that are written in other languages, but they are not perfect, like this example.

Facebook recently translated a Colombian voter's post from the original Spanish to English. It reads, (reading) life gave me the opportunity to vote in my country on this day - so exciting. It's not poetry, but you get the idea. Now Google says it can do it better thanks to artificial intelligence. NPR's Aarti Shahani is with us to talk about this. Hi there.

AARTI SHAHANI, BYLINE: Hi.

MCEVERS: So how is Google using artificial intelligence to improve translation?

SHAHANI: Well, you know, so it's Google as well as Facebook, Microsoft, Baidu and a bunch of other tech companies, leading tech companies who are moving to an AI system here. And this is how it works in principle. You could have a linguist, a human expert, sit down and dictate all the rules going from one language to another, but that would be a really tedious, exhausting process that arguably, as it seems, doesn't really turn out the kinds of results you want because a lot of languages aren't structured in parallel form.

So what these companies such as Google are doing now is they're moving to a pattern recognition. It's called Neural Machine Translation. And the way it works is, you take a boatload of data - translations - really good translations, for example, between Chinese and English. You take that data, and you load it into the computers. And the algorithms then mine through the data to look for patterns. Oh, this seems to go to that. That seems to go to that. And so by doing this sort of mining through and pattern recognition, the machines figure out how to translate not just phrase by phrase but entire thoughts, sentences, paragraphs.

MCEVERS: We wanted to see how this method compares to a professional translator.

DOTTIE LI: My name is Dottie Li. I'm managing director of TransPacific Communications. I happen to be the voice and voice coach of Rosetta Stone's Mandarin products.

MCEVERS: Dottie Li was kind enough to go head to head with the new Google Translate for us. We gave her this commentary about the U.S. presidential election from People's Daily.

UNIDENTIFIED WOMAN: (Speaking foreign language).

MCEVERS: This is how Google translates it.

COMPUTER-GENERATED VOICE: Everyone is most concerned about Hillary and Trump, who is even worse for China, when this problem is not clear to see them both who may destroy the United States.

MCEVERS: And this is how Dottie Li translates it.

LI: What people are most concerned with between Hillary Clinton and Donald Trump is which one of them would be tougher on China. When they are unclear on the answer, they would have to see which one would possibly make America worse.

MCEVERS: So we asked Dottie Li what she thought of all this.

LI: I think this is definitely a battle between human intelligence and mechanical robotic computer devices. I firmly believe that human beings will triumph. In this case, you could see that machines came through with some kind of mechanical, very technical aspects that don't make much sense.

MCEVERS: What about you, Aarti Shahani? What do you think about all this?

SHAHANI: (Laughter) Well, I'm thinking two things. One is I'm probably a bit more optimistic than she is about the capacity of machines to translate more beautifully. And two, maybe more importantly, is, you know, she used the word battle. This is a battle. And I think that the people engineering these kinds of translation techniques see it more as a partnership.

You know, when you look at, for example, Google and Google's take on its own work right now, they recently issued a paper on this. And in translation, like, on a scale of zero to six - zero being total garbage and six being perfect - humans in Chinese-to-English, English-to-Chinese will get about a five. The old way of translating phrase by phrase would be around, like, a three and a half, three and three quarters. This new way of using neural nets to do it gets you about a four-and-a-half.

So you know, by Google's admission, not quite human but better than what was before - you know, and I think that part of their hope in releasing this Chinese-English translation out, you know, to the public in the wild before other languages get released out there is because they recognize, man, the old way of doing it was so bad. They needed some sort of improvement for the public.

MCEVERS: That's NPR's Aarti Shahani. Thanks a lot.

SHAHANI: Thank you.

Copyright © 2016 NPR. All rights reserved. Visit our website terms of use and permissions pages at www.npr.org for further information.

NPR transcripts are created on a rush deadline by Verb8tm, Inc., an NPR contractor, and produced using a proprietary transcription process developed with NPR. This text may not be in its final form and may be updated or revised in the future. Accuracy and availability may vary. The authoritative record of NPR’s programming is the audio record.