Remodel The Net's Tower Of BabelRemodel The Net's Tower Of Babel

But getting Web sites and the back-end databases that drive them set up to provide catalogs or other documents in the language the customer wants to use takes some work. The SDL system has a 15,000-word vocabulary, but users have to add words it doesn't know; the software supplies a list of those words after scanning a company's documents. Some of the most challenging work involves preparing the software to deal with a particular industry's jargon.

"A classic example might be the word 'finish,'" C.R. Laurence's Roach says. "In normal, everyday English, 'finish' typically means 'to end.' But in my industry, the word 'finish' most often represents a color." So Roach had to load some new rules and custom phrases into the software's dictionary. It put in a translation that indicates the words "brass" and "finish" used together are related to color, but the phrase "almost finished" doesn't mean color.

The project has been so successful that C.R. Laurence is planning to add more languages, starting with Italian. Complaints about poor translations have been rare, but optimization of the system remains a maintenance task. "That was and is an ongoing process," Roach says.

The New York State Department of Motor Vehicles faces similar challenges. It's beginning to translate customer-service information on its Web site into 20 languages to serve New York's diverse population, starting with Spanish. After 11 months, out of 333 documents, only 4% have been through the process. That's because the department is taking a thorough approach to assure accuracy, says Richard Vang, a project assistant in the department's Internet services office.

Vang and his team members, who also use the SDL software, are first standardizing each page into plain English versus bureaucrat-speak. Using the MAXit Controlled English Checker from Smart Communications Inc., they're painstakingly going through page after page on the site, making sure terminology is standardized, text is in active tense, and so on. It takes three or four days to complete an average-sized document, and after it's translated it goes through two levels of checking by employees who are native speakers. For a typical Web page, Vang figures it could take six months for a document to get all the way through the process.

The cost of the project, which the department declined to disclose, is justified, Vang says, by the 90% to 95% accuracy rate of translations. In a sense, the process is even better than using human translators, he says, since new pages can be translated in real time, thanks to all the work the department has done to build a smarter system and a detailed custom dictionary. That means a page that hasn't been reviewed or standardized can still be translated with 75% accuracy. In these cost-constrained times, that beats the alternative. To actually hire somebody to translate full time for each language wouldn't be very cost effective, Vang says.

Businesses might not even need bilingual employees to talk to customers in real time, either online or in the real world. Enterprise translation engines are already capable of instantly translating con- versations in a chat room, says Richard Kreidler, who advises clients on machine-translation projects for the Mesa Group. In the months and years to come, the software is even likely to power handheld devices that can interpret speech. The U.S. Department of Defense is already developing such a project, he says.

"One of the first steps in globalization is to remove the language barriers," Kreidler says. As more consumers are getting onto the Net, "people are becoming empowered. They're saying, 'This is not acceptable. Talk to me in my language. Respect me, and I'll respect you.'"