Conversing with Klingons may be a piece of cake in the Star Trek universe, but here on earth, language translation apps still have a long way to go. Two Silicon Valley teams, at Google and SRI International, are making good progress, but will they ever get close to the holy grail: translations as accurate and reliable as multilingual humans? Reporter: Alison van Diggelen
Conversing with Klingons may be a piece of cake in the Star Trek universe, but here on earth, language translation apps still have a long way to go.
In the race to improve machine language translation, two Silicon Valley teams, at Google and SRI International, are making good progress, but will they ever get close to the holy grail: translations as accurate and reliable as multilingual humans?
At the Google Headquarters in Silicon Valley, a team has been working since 2001 on Google Translate, software that aims to translate the entire world’s information. Senior communications associate Roya Soleimani demonstrates how she can access web pages in over 70 languages with the click of a mouse. She can also get instant translations on her smart phone of text, speech, and even photos that she’s taken (on her phone) of menus and street signs.
Dreams of Star Trek
“You can have access to the world’s languages right in your pocket,” says Soleimani. “The goal is to become that ultimate Star Trek computer.”
She explains how instant webpage translation has become an indispensible tool for California school districts and cities to reach their multicultural communities.
“Google’s mission is to organize the world’s information, make it universally accessible and useful,” adds Josh Estelle, who leads the Google translate team for mobile.
So, how exactly does Google do it? Not Bilingual Elves
“Right so it’s not bilingual elves, as much as people might think,” laughs Estelle. “Our system is built off example data…we use what’s called statistical machine translation.”
Remarkably, there’s not one linguist on Google’s core team: they’re all engineers or statisticians, using computers to analyze billions of translations every day. But as our demo in Google’s Yoshka’s café shows, it’s not there yet. Estelle speaks into his android phone.
“Let’s try something in Spanish…Do you listen to KQED?” he asks.
And the smartphone replies, “Que escuchas K?”
Close, but no cigar.
We try a few more phrases with Seth Rosenblatt, a senior writer at tech product reviewers, CNET in San Francisco. He types into his laptop: “I fancy my girlfriend.”
And a big smile spreads across his face, as he chuckles, “In Japanese… probably not good to say on radio.”
There are countless examples of translation bloopers by Google and other competing services. According to Rosenblatt, these include the Bing Translator from Microsoft, Babylon software from Israel, and an iphone/android app called iTranslate from a team in Austria.
So why is it so hard to get it right?
“It’s a very deep computer science problem…there are so many idioms, nuances within language…different morphologies,” explains Soleimani.
Not to speak of dialects, accents and ambiguities. Spanish words like “vieja” (old lady) or “gordito” (little plump one) can have several shades of meaning -- from insults to terms of endearment, depending on the context.
The Google Way
So what’s the solution? The Google team tackles it “the Google way.”
“We can use those mathematical algorithms…to process all this data and make sure we pull out the nuggets of truth, right…the good pieces of translation,” says Estelle.
So Google translations are only as good as the translated data it finds online. Sources like the United Nations and European Community translations are reliable, but these only make up a small fraction of translation data online.
Language researchers at Silicon Valley’s SRI International, inventors of Apple’s Siri, have a different approach that’s more mission-specific.
The translation lab’s director is Kristin Precoda and her team demonstrates a smartphone app they developed for military personnel in Afghanistan to translate between English and Pashto.
Colleen Richey speaks the Afghanistan language Pashto into an android phone equipped. She says, “zmA mAshum mariz day.”
And the phone, after a few seconds, speaks back in English: “My son is sick.”
Her colleague Harry Bratt then replies in English: “I can take you to a hospital.”
And the phone translates that into Pashto.
Asking For Directions
Precoda says that her team is working on a cutting edge strategy , one we might call the enlightened female approach, since it programs computers to ask directions.
Her team is equipping computers to learn when they’re making mistakes, and ask for clarification when they need it.
“In the demo we just showed you where it’s interactive translation between two people. Those people are smart, leverage them,” explains Precoda. ‘It’s definitely a step forward, this whole idea of asking the user: help me out so I can serve you better, give you better translation.”
The Holy Grail
So what is the holy grail of translation? For Precoda it’s as simple as the Hippocratic oath: “First do no harm.”
“In military situations, medical situations,” she says, “A poor translation can really do a lot of damage.”
So will computers ever be as reliable as bilingual humans? Jill Pellettieri is an associate Spanish professor at Santa Clara University in San Jose. She thinks it will never happen.
“I would use the analogy of a supermarket and the self checkout lane,” says Pellettieri. “There always has to be someone there watching over machines because they make mistakes.”
But Google is in a rush to win the translation race, even if that means mistakes.
“We err on the side of launching things as quickly as we can, as early as we can,” says Google’s Josh Estelle. “We dream of that Star Trek Computer…we think we can eventually get there.”
The Role of Google Glass
CNET’s Seth Rosenblatt agrees. “Google Glass plus Google Translate could lead to something like a Star Trek Translator...”
But, he cautions, not any time soon. For now, SRI’s Precoda has this advice for using online translations: “Trust, but verify.”
Find out more about translation:
World Reference Forum -- Online resource to check translation accuracy with native speakers
SRIs Speech and Technology -- Find out about the STAR Research Lab and the language modeling toolkit.
SRI’s BOLT -- program with DARPA: Language translation for military personnel in the field