Re: Ontological EDI & Wittgenstein
"Peter Clark" <pclark@cs.utexas.edu>
Message-id: <199410031406.JAA20332@sage.cs.utexas.edu>
Subject: Re: Ontological EDI & Wittgenstein
To: fritz@rodin.wustl.edu
Date: Mon, 3 Oct 1994 09:06:13 -0500 (CDT)
From: "Peter Clark" <pclark@cs.utexas.edu>
Cc: edi-new@tegsun.harvard.edu, srkb@cs.umbc.edu
X-Mailer: ELM [version 2.4 PL24alpha3]
Mime-Version: 1.0
Content-Type: text/plain; charset=US-ASCII
Content-Transfer-Encoding: 7bit
Content-Length: 2192
Sender: owner-srkb@cs.umbc.edu
Precedence: bulk
> (Fritz Lehmann) A human being has to read the documentation and
> address questions like "Does 'vehicle' for system A include
> aircraft, as in system B, and if not, how can we use B's database
> of vehicles for the 'vehicle' records needed by A?".
and does 'aircraft' include 'helicopter'? and does 'helicopter' include
'tilt rotor planes'? etc.
Similarly:
>>> (Fritz Lehmann) "we can agree completely that no barge ... is a suburb"
>> (Scott Dickson) Or, we might not agree. The question is, do we have to?
> (Pat Hayes) Yes!! That is, we (and our programs) have to somehow come to
> sufficiently close terms of agreement in order to be able to communicate
> properly.
What ultimately "grounds" this vocabulary -- what are the primitives that
we can be sure we agree on, from which other like "barge" and "suburb"
become defined? Maybe they are simply examples. I know my Nissan Sentra
and my friends Dodge Colt are cars. I know my desk and chair aren't.
There's probably some borderline cases in between, but do we have to
agree on these if they never occur in practice?
Some similarities with inductive learning theory comes to mind -- namely
that you can *never* be sure you and someone else agree completely on
a concept definition, only that you agree on an ovewhelming fraction
of commonly occurring cases. This is the basis of the PAC learning
framework, and a lot of other inductive learning work. Valiant
says ("A Theory of the Learnable" Comm ACM (27) 11, 1984):
"If (such) a [learnable] concept is passed on among a population in a
distributed manner, substantial variations in meaning may arise. More
importantly, what consensus there is will only be meaningful for natural
inputs. The behaviour of an individual's program [for identifying concepts]
for unnatural inputs has no relevance. Hence thought experiments and
logical arguments involving unnatural hypothetical situations may
be meaningless activities."
--
Peter Clark Dept of Computer Science
email: pclark@cs.utexas.edu University of Texas at Austin
phone: (512) 471-9574 Austin, Texas, 78712
fax: (512) 471-8885 USA