Writing by Peter Hilton

Machine Translation Of The World Wide Web

The killer application for machine translation.

I have just read Machine Translation, An Introductory Guide, which describes the use of computer software to translate between human languages. Like all the best things the IT industry will sell you, this is technology that has been around for years. The book mentions METEO, a system used to translate weather reports from English into French and says 'The METEO system has been in daily use since 1977. As of 1990, it was regularly translating around 45 000 words daily.'

Café des Tramways, Luxembourg Nowadays low-cost software is available for desktop PCs that will make a fairly good job of translating between a selection of languages.

In the future we can expect MT systems that perform instantaneous speech-to-speech translation so that we can communicate in a foreign language over the telephone or even, with smaller hardware, via hearing-aid style devices. Actually, the necessary machine translation, continuous speech recognition and speech synthesis technologies are probably now fairly close to being able to provide this.

The Killer Application For Machine Translation

It is clearly a truism to say that for the World Wide Web (WWW) to be truly world-wide its content must be accessible to people all over the world. The problem for many people around the world is that the majority of the content is in the English language. Web browsers that could automatically translate Web pages into the user's local language would have two great benefits.

Firstly, if (or when) the WWW becomes as pervasive a medium as television, being able to read Web pages in one's native language would do a lot to protect minority languages from the 'anglophone cultural imperialism' that the French are so worried about. Secondly, it would give anglophones access to other cultures; something that people usually start to appreciate when they learn one or more foreign languages and discover that it is not just the words that are different.

On the WWW this could work in two ways. Firstly, Web browsers would be able to translate from half a dozen major languages into the user's native language on the fly. This would allow people in Kazakhstan, say, to read most of the world's Web pages in the Kazakh language. Secondly, if you accessed a page on a Web server in Western Java, say, the server would automatically translate this from Sudanese into one of half a dozen other languages. The combination of these two approaches would probably allow most of the world's population to read most of the world's Web pages.

To a certain extent all of this is already possible, thanks to applications such as Digital's Babelfish page, which translates between English and a few other European languages. This greatly opens up the WWW to native speakers of French, Spanish, Portuguese, German, and Italian. I am interested to see whether any non-European languages will be added.

It is reassuring so see technology helping support cultural diversity in this way; I am looking forward to using Babelfish to read some foreign newspapers that I have never read before. Unfortunately too many people still say, as did the American in his twenties who I met at the weekend, 'I don't know why everyone doesn't just learn to speak English'. He was less than half joking. I can only hope that his children do well in their Mandarin Chinese and Hindi lessons at school.

Machine Translation, An Introductory Guide - D. Arnold, L. Balkan, R. Lee Humphreys, S. Meijer, L. Sadler. ISBN 1-85554-217-X, out of print

Comments

After spending about 15 years on creating a translation program for PCs I can tell you this: We aren't ready yet! At least not for translating any kind of text covering any kind of topic. A machine must first know what kind of topic to translate. Some words have many different meanings, depending on the subject matter.

For example a simple word with three letters like 'bar' can mean a cocktail bar, a handle bar on a bicycle, a space bar on a keyboard, an exercise bar for gymnastics, a sand bar in the ocean, a bar of a popular tune, a bar for lawyers, a salad bar. And the list goes on. A machine can only produce a correct translation when it 'knows' how to select the right bar.

We understand how this information can be embedded in a program (at least I do), but nobody has yet been willing to come up with the funds. Until that happens, machine translation will remain a very primitive type of application for computers.

Contributed by Ralph Dessau on 3 November 1999.

Sure, perfect machine translation is a 'hard' artificial intelligence problem that we're not going to solve any time soon. I don't think that that's a particularly helpful objection though, because translation doesn't have to be perfect to be useful - if that were the case, most human translators would be out of a job.

You could even say the same thing about spell-checkers - they can only 'guess' at what is a spelling error. That doesn't stop imperfect machine translators being useful. I even find babelfish.altavista.com useful.

For example, I got an e-mail in German a short while ago, which I couldn't understand. However, Babelfish's imperfect translation was good enough for me to be able to understand the main point of the mail, and allowed me to reply.

It's like the example of the eye in human evolution. The creationist crowd dispute that something so complex could spontaneously 'evolve'. What they've missed is that a basic proto-eye is more useful than nothing at all.

Contributed by Peter Hilton on 3 November 1999.

Everything you say is true. However, my comments are based on all the negative comments I have received over the years from ignorant people, who refused to understand what you obviously have figured out. Such people expect perfect output from a machine without a brain. Obviously, in many cases, their own brains weren't that great either.

In case you are interested, you should look at our website: http://www.pctranslator.com/html/multi-language.htm. We outperformed all other participants in an ARPA MT gathering some years ago. I personally developed our dictionaries in 18 different language pairs. Thanks for your response,

Contributed by Ralph Dessau on 3 November 1999.