Re: Universal Translation Language

On Fri, 28 May 1999 06:53:37 +0100, "Raymond A. Brown"
<raybrown@...> skribis:

Hi, Raymond, how are you?

>>Here's the thing.  Wittgenstein's corpus is very much concerned with =
the
>>question of ambiguity, and ultimately with whether or not it is a
>>universal good to eliminate ambiguity.
>
>Indeed - in fact such a language would make some things quite
>untranslatable since the effect the writer is trying to produce depends
>upon a certain degree of ambiguity.  And this varies considerable from
>language to language and is one of those things that cause headaches to
>translators.
One thing that is clear is that on any translation (either manual or
automatic) it's going to be some loss (or undesirable addition) of
information. But that is not going to prevent us from doing
translations, cause they are a need.

>Indeed, I have grave doubts whether it is possible to eliminate all
>ambiguity.  Classical Yiklamu has possibly the largest vocabulary of all
>conlangs and was design specifically to be as unambiguous as possible as
>regards lexicon.  But I suspect one skilled in its use could still be
>umbiguous if s/he wished to be.
Indeed, totally elimination of ambiguity is probably impossible. But
that's not really important if we're talking about "enough
unambiguity" rather than "total unambiguity", as is the case. "Enough
unambiguity" means that the m-translated text will be closer enough to
author's ideas so it gets worthy. However, I wouldn't expect generated
text to be 100% correct syntactically nor semanthically. Some human
post-editing would be necessary if quality is a determining factor,
but it'd be quite little, so publishing companies will may take great
advantage of it.

>The whole matter of translation is full of pitfalls.  Another book worth
>reading in this regard is George Steiner's "After Babel".  I have grave
>doubts about the possibility of universal automatic MT.
That depends on how you define it. NL to NL MT is a process which we
can divide in two parts: parsing of source NL and generation of target
NL. The former is the hardest, and is the source of most errors,
because of NL inherent analisys problems for a machine. But most of
those problems may be surpassed if source language is a MT-oriented
conlang.

>However, I do wish Marcus luck with his efforts.  Be aware of the =
pitfalls
>and even if (as frankly I suspect) you will not be successful in =
achieving
>all you wish, you may well make some interesting discoveries along the =
way.

I cannot determine with exactitude the quality expectable from a UTL
-> NL machine translation, but no doubt it'll be quite better than a
NL -> NL one. Just for that reason I know that UTL may be of high
utility for multilingual text generation.



Saludos,
Marcos