Automatic Translation: Promise or Pipe Dream?

Countless Web sites and software packages promise translation at the touch of a button. How useful are these tools to healthcare packaging professionals?

by Robert Sprung

"Can't you get a machine to translate it?" Given computing power these days, the question seems well worth asking. Do the math: a company with 100 instructions for use totaling 100,000 words in English contemplates a move into 10 foreign languages; the price tag for translation could easily top $300,000. If the money isn't an issue, consider the potential time savings. Translating those 100,000 words with humans and word-processing equipment could take at least two to three months. With the stakes this high, and with the price of computer processing and storage plummeting, pressures to automate translation are considerable.

First-time buyers are often unaware that automatic or machine translation has been the Holy Grail of AI (artificial intelligence) for much of the last half century. A machine that can translate is a machine that can think and write like a person. Easier said than done, it turned out. Take the sentence "Mary had a little lamb." How do we know that had means "possessed" and not "ate" (as in, "John had a little beef and Mary had a little lamb"), "took advantage of," or, heaven forbid, "gave birth to?" The answer is, we don't know—cognitive scientists do not yet sufficiently understand how the human brain places this particular lamb in a nursery rhyme and not in a butcher shop. And if the computer has a problem with Mary and her lamb, what sorts of problems will it have with caution statements or product warranties for a medical device or a drug?

Although commercial machine-translation systems have been around for about 30 years, the vast majority of translation performed today is still done by humans. Douglas Hofstadter, a pioneer in AI, thinks that the computer may never be able to produce a reliable translation that reads coherently.

Which is not to say that machine translation, or MT, does not have practical uses. Some think the solution lies on the output side, through lowered user expectations. AltaVista pioneered a Web-based MT solution by partnering with Systran. Click on a Web address listed in the search engine, select a language, and voilà: a "translated" Web page free of charge. Don't expect a polished translation, however. The system is good for "gisting"; by having most key words translated accurately with reasonably understandable syntax, you can often grasp enough to determine basic meaning or at least context.

Another angle is to refine the input. If you can simplify the text to be translated, the computer's task will be greatly simplified. Caterpillar, justly concerned with the cost and time of translation, helped develop the concept of "controlled authoring." By dramatically limiting the vocabulary and sentence complexity of source text, the company greatly improved the performance of machine translation. Some sophisticated users of controlled language have authoring tools that compel technical writers to use very simple sentence structure with limited vocabulary. While this boosts accuracy, it may take years to implement such a system.

Here are some pointers in setting realistic expectations for MT and determining whether it can work for you:

1. Prepare for a large-scale investment, in both time and money.

2. Expect a long implementation cycle. It may take several years to build the dictionaries necessary for accurate MT.

3. Perform a professional cost-benefit analysis. Machine translation generally makes sense only if you have very large volume on a recurring basis in a relatively narrow domain.

4. Prepare to deal with internal resistance among your in-country reviewers, who may not be open to the use of MT. First-pass quality is always lower with MT than with a human.

5. Pick a stable vendor with strong reference accounts. To learn more about MT, contact the Association for Machine Translation in the Americas at AMTAinfo@att.net.

Robert Sprung is the chairman of Harvard Translations (Cambridge, MA). His latest book, Translating Into Success, includes case studies in multilingual labeling and is available on amazon.com. To contact him, send e-mail to robert_sprung@htrans.com or visit www.htrans.com.

No votes yet