Adam sent me this article about the latest developments in machine translation.
It appears that some clever researchers have managed to write software which "learns" languages using statistical analysis and probability techniques, (presumably from a fantastically large corpus!)
Uh-oh... here we go again. If there's one activity which is purely based on human intelligence, it's translation (and interpreting, for that matter).
Good translation is not really about swapping words, and it's certainly not about rules of probability. What it is, however, is very simple.
It's the following process:-
1) Read and understand a text
2) Render that text in a way that is understandable for the new audience.
Sound simple? Well, it's really that word "understand" which is the clincher. It's a fractal pattern of complexity, all about context. Who wrote the text? What was the intended audience? Is there another "hidden" audience? Why that word? Is it a cultural or intertextual reference?
I know that all sounds really... well... pretentious, but just try using any of the currently available machine translation tools and see how quickly they get tripped up.
Understanding is the one thing which I for one am quite happy that machines can't do. And until we fully understand consciousness ourselves, I don't think we can ever bestow it.
But consciousness, understanding and intelligence can be quite well mimicked. I'm sure you've heard of those experiments where humans are challenged to type into a terminal and see whether their correspondent is another human or a computer. (I even had a piece of software chatting to me on MSN a few months ago).
But it's all tricks, snake-oil, sham. Who am I to say it, but I remain an AI sceptic. To me it looks like people who are trying to achieve space travel by finding ever taller trees to climb.
Posted by Eurodan at February 22, 2005 8:16 PM