From newshub.ccs.yorku.ca!ists!helios.physics.utoronto.ca!news-server.csri.toronto.edu!rpi!usc!cs.utexas.edu!uunet!psinntp!scylla!daryl Thu Dec 26 23:58:02 EST 1991
Article 2357 of comp.ai.philosophy:
Newsgroups: comp.ai.philosophy
Path: newshub.ccs.yorku.ca!ists!helios.physics.utoronto.ca!news-server.csri.toronto.edu!rpi!usc!cs.utexas.edu!uunet!psinntp!scylla!daryl
>From: daryl@oracorp.com
Subject: Re: Machine Translation
Message-ID: <1991Dec21.230721.13166@oracorp.com>
Organization: ORA Corporation
Date: Sat, 21 Dec 1991 23:07:21 GMT

Mikhail Zeleny writes:

>(Robert Kohout) writes:
>>       [...]              if someone in this debate sees how
>>the correctness of Searle's position in any way implies that,
>>for example, we will never be able to engineer a fully automatic,
>>high quality machine translator I wish they'd explain it
>
>Simple.  Correct translation is a matter of finding an approximate
>synonym; synonymy is a semantic relation; if machines can't compute
>semantic relations, they can't translate anything.

Mikhail, I don't think your argument is correct. If you place a
reasonable bound on the length of the input string (say, ten trillion
words) and a similar bound on the length of the output string, then
the perfect translator on such a domain would be a function (or a
relation, if you want to allow nondeterminism in the output) from a
finite domain to a finite range. All such functions are computable by
finite state machines.

If there is any sense in which natural language translation is
inherently noncomputable, it would only show up as needing a new
translator when you have the first translation task for texts of
longer than ten trillion words.

For these reasons, even if human minds were essentially
non-algorithmic, that non-algorithmic nature cannot possibly make a
difference if you ar only considering bounded-length conversations
(and, so far, conversations have been bounded by what can be said in
120 years or so).

Therefore, any argument (such as Penrose', which as you know, I
consider to be baloney) that purports to show that humans are capable
of something non-algorithmic cannot possibly make an *observable*
difference for humans with finite life-spans if the output is forced
to be discrete (a sequence of words, or proved theorems, etc.).

Daryl McCullough
ORA Corp.
Ithaca, NY



