How good is machine translation?

29 September 2005

Download this program (Right click and Save As...)

When the computer was invented, one of the first things people thought about was how you might use it to translate foreign languages. But early efforts at machine translation... fizzled. And in the late 60s and early 70s the effort was almost completely abandoned. You've probably heard the funny mis-translation stories: "The spirit is willing but the flesh is weak" translated as the "The whisky is strong but the meat is rotten." Or the term for a device known as a "hydraulic ram" being translated as "water goat." Machines can indeed translate languages, but the Holy Grail for humans who use them -- "fully automated high quality machine translation" -- is still elusive. What's the problem here?

Well, the answer lies in the complexity of language in general. It begins at the word level. Take the word "bark": a computer doesn't know that a dog isn't covered with wood, or that a tree doesn’t make loud noises when someone approaches. Which translation of "bark" should it choose? The Spanish verb "comer" might be translated as "eat, capture, overlook, corrode, fade, itch, skip, slur, swallow, or take”, only one of which might be correct for a given sentence.

At the sentence level it gets even worse. Take this sentence: "John saw the woman in the park with a telescope". There are about five or six possible meanings for that sentence -- for starters, who do you think has the telescope, John or the woman in the park? Or is it John who's in the park? The mind boggles.

The less one language resembles another, the harder the problems of translation are. Where English has three words, "he", "she" or "it", Turkish has just one word, "on". Lacking any cues, how do you translate "on" into English? Russian has two words for "blue" where English has only one -- so if you want to translate "She's wearing a blue dress" into Russian, you have a 50% chance of choosing the right word. There are many cases in which you just can't make an exact translation between languages.

And now for the good news. Despite its limitations, there are situations in which the machine does a very good job, particularly if the topic of translation is narrow. The Canadian Meteorological Centre, for example, uses machine translation for bilingual weather reporting. In the commercial world, technical writers have learned to write manuals and parts catalogs so that computer-based translations -- let's say from Japanese to English --need only a little correction. Even on the Internet, the click of a mouse can translate entire web pages with fairly good quality, at least for some languages.

And let's remember that machine translation doesn't always have to be perfect. It can be very useful even if it has some garbles. For instance, where human translators are scarce and material to be translated is great -- as frequently happens in the post-9/11 world -- machines are widely used to "triage" material before it reaches the desk of an overworked human. The National Security Agency's "CyberTrans" software, created for the Department of Defense and the Intelligence Community, does precisely that: it identifies a language, corrects misspelled words, and translates some sixty-five languages into English, with quality that's usually good enough for scanning.

And the future looks good. Experts are using new approaches to machine translation, with encouraging results. The quality of Arabic-to-English translation, for example, has improved dramatically in the past five years. Machines are making multilingual chat rooms possible on the web and helping multinational military forces talk to one another. Yes, a human translator is better. But even though the goal of "fully automated high quality machine translation" is still not in sight, we've come a long way. Steady progress is being made in quality, and the use of machine translation is increasing by leaps and bounds. The machine... is here to stay.

That linguistic thought comes from Dr. David Savignac, Director of the Center for Applied Machine Translation at the National Security Agency. And this is the Five Minute Linguist at the College of Charleston, in cooperation with the National Museum of Language. Visit us at www.cofc.edu/linguist . And keep in mind that wherever you are, and whatever you do... language makes a difference.