Re: natural language translation (was Re: How factual are second-hand translations?)

From: Jacques Du Pasquier (
Date: Wed Jan 09 2002 - 12:52:16 MST

I learned programming with a computational linguist, and I had
occasion to get an idea of the field.

The (somewhat intuitive) conclusion I got from it is that to build an
adequate general-purpose translator, you need to build and AI that
actually understands the text (i.e. a human-level AI, with common
sense and everything). Otherwise it won't work, except in restricted
domains (say, if the program can assume that you only feed him weather

So we will probably have that within a few decades (through uploading
if not programming) -- but probably not before. Of course you can have
tools that help you understand a text by looking up the words in a
dictionnary, and doing a bit better than that, too, with some kind of
syntactic parsing and synthesis, various heuristics to establish
context, exception lists and the like. But telephone translation, for
example, being spontaneous talk by human beings, needs general-purpose
translation, and so it won't work before we have human-level AI.


Dossy wrote (9.1.2002/09:14) :
> On 2002.01.08, Spike Jones <> wrote:
> > Suppose we had a list of special cases that the software
> > would look for in common speech, then translate as special
> > cases. A few examples I can think of would be
> >
> > "...six of one, half dozen of the other" = identical.
> > "...wassup?" = hello
> > "...have a nice day" = goodbye, get outta my face, etc.
> >
> > Computers are now fast enough to tear thru a jillion of these
> > kinds of special phrases that might trip up the literal translators.
> Also, consider the traditional problem of translating idioms.
> Things like "kick the bucket" ... a literal translation of

I learned the

> "Last night, poor John kicked the bucket."
> might yield something about a poor man named John kicking
> a pail the previous night.
> Things like Ebonics and other dialects of language show us
> exactly how impossible natural language translation is, unless
> we build a generic "system" that can learn languages. There's
> no value in building a "database" of translations because it'll
> never be able to keep up with the languages (unless they're dead
> and no longer spoken or changing).
> The big win in translation will be semantic translators, ones
> that derive the meaning of a passage (not just a word or sentence,
> but whole phrases at a time within the context of a larger passage)
> and then translate it into the equivalent *meaning* in the target
> language following all the rules and structure of the target
> language.
> -- Dossy
> --
> Dossy Shiobara mail:
> Panoptic Computer Network web:
> "He realized the fastest way to change is to laugh at your own
> folly -- then you can let go and quickly move on." (p. 70)

This archive was generated by hypermail 2.1.5 : Fri Nov 01 2002 - 13:37:33 MST