Translation by Numbers
Bryce Benjamin knew a winner when he saw one. It was December 2001, and the infotech entrepreneur was meeting with two professors who were starting a company to commercialize “statistical machine translation.” Their breakthrough: software that could learn automatically to translate text from one language into another.
Benjamin believed the technology was desperately needed for purposes of both homeland security and business communication. The company’s location – on the water in sunny Marina del Rey, CA – didn’t hurt either. “I looked out at the view,” he says, “and I thought, ‘This deal has a lot of promise.’”
Today the trio’s 35-person startup, Language Weaver, is one of the leading companies in the burgeoning field of machine translation. For U.S. counterterror translators facing a growing backlog of untranslated audiotapes and communiques, software is increasingly the weapon of choice. Multinational corporations like Google, Yahoo, and Microsoft – not to mention smaller companies with global staff – are also driving the demand for machine translation of technical documents and Web pages.
Language Weaver’s software translates text between English and half a dozen other languages, including Arabic, Chinese, and Spanish. So far, the technology is most useful as a screening tool that monitors reams of foreign-language news broadcasts, chat rooms, and websites. “People use our translation software to determine the relevance of information, as a triage function,” says Benjamin, the company’s CEO. “It’s very good at telling what a certain passage is about.”
Most machine-translation systems work on individual words or use complicated sets of translation guidelines, which must be devised by linguists and coded by hand. Language Weaver’s technology, which company cofounders Kevin Knight and Daniel Marcu developed at the University of Southern California’s Information Sciences Institute (ISI), takes a different tack. It uses human translation data, such as United Nations transcripts, to set up “parallel corpora” of text passages in two languages, aligned sentence by sentence.
From these side-by-side comparisons, the software learns to translate between the languages – extracting statistical patterns that indicate that a particular grouping of words in Arabic, say, tends to correspond to certain words in English. The system translates phrase by phrase, so if it encounters the words “interest rate,” it will associate them with banks and finance, not curiosity and speed. And the machine-learning approach means the translations should improve with time. “The more data you add, the better the performance will be,” says Knight.
Until recently, this statistical approach, which has roots in 1940s wartime cryptography, was too slow to be useful. But on a modern PC, Language Weaver’s software can translate 5,000 words a minute at state-of-the-art accuracy levels; on a network of servers, it can handle up to 500,000 words a minute.
The names of the company’s U.S. government customers are tightly under wraps, but Benjamin says feedback from the intelligence community has been overwhelmingly enthusiastic and that the technology “played an important role in a mission that saved lives.”
But human translators won’t lose their jobs anytime soon. To accurately do some translations – of, say, technical material – Language Weaver’s system must be trained on similar texts translated by hand. And some experts are skeptical, since the statistical approach hasn’t solved the deeper problem of getting computers to understand natural language.
“It is good for the field that companies like Kevin’s succeed,” says Sergei Nirenburg, a machine translation expert at the University of Maryland, Baltimore County. But, he adds, “you will still have to file the product under ‘Good Uses for Crummy Machine Translation.’”
Nevertheless, Language Weaver has been profitable since late 2003 and is now investigating business applications. At ISI, Knight and Marcu are testing software for a handheld translator that can handle questions and answers between doctors and patients. An outside company is also developing a real-time Arabic-English translator for instant messaging built around the Language Weaver software.
But the biggest market may be multilingual search: typing a query in English could bring up scores of previously invisible foreign websites that could be translated into English. In five years, says Knight, expect to be able to read any Web page in practically any language. “You’ll wonder how you ever did without it,” he says. “But you’ll still laugh at the translation mistakes.”
Geoffrey Hinton tells us why he’s now scared of the tech he helped build
“I have suddenly switched my views on whether these things are going to be more intelligent than us.”
Meet the people who use Notion to plan their whole lives
The workplace tool’s appeal extends far beyond organizing work projects. Many users find it’s just as useful for managing their free time.
Learning to code isn’t enough
Historically, learn-to-code efforts have provided opportunities for the few, but new efforts are aiming to be inclusive.
Deep learning pioneer Geoffrey Hinton has quit Google
Hinton will be speaking at EmTech Digital on Wednesday.
Get the latest updates from
MIT Technology Review
Discover special offers, top stories, upcoming events, and more.