On Tue, 31 Jul 2001, Gregory Himes wrote:

> > Does anybody know what the main algorithm used for
> > translation program?  Does it look up each word,
> > phrases or whole sentences?
>
> I know how they work in general.  It must first look up every word in a
> dictionary, then it must assign a code to each to determine whether it
> is a noun, verb, etc.

And then it is already fucked. What spoils this strategy and makes machine
translation unfeasible for the foreseeable future is that very often there
is no 1:1 mapping between words in two languages. A Japanese word may have
two slightly different applications which require different English words
to translate (which will often have other nuances not present in the
Japanese word). Which one is the corect choice depends on context. And that's
just the most simple example. In general, natural languages are
*extremely* context sensitive, and that absolutely kills translations
programs, because to build a sentence, they must look at the context, but
they can't get a context analysis running without having a complete
sentence. An iterative approach might be the solution, but I suspect it
would often just amplify mistakes.

For an example, look at http://www.tashian.com/multibabel/

It may be somewhat unfair, but you'll find that it horribly mangles even
the most simple sentences. Try "This is a pen"...



Michael "Brazil" Borgwardt --- Member of #WASHU# and Her would-be guinea-pig.
|   Untiring defender of Washu-chan, Asuka-chan and Elektra-chan.           |
|       A Homepage for Elektra: http://brazils-animeland.de/elektra         |
|           Tenchi Fanfiction Reviews: http://brazils-animeland.de/tmffr    |
+-------------- Let`s shake the dew off this lily, shall we ? --------------+