Re: Converses (Re: Availability of the ANSI standard proposal?)
Peter Clark <pclark@cs.utexas.edu>
From: Peter Clark <pclark@cs.utexas.edu>
Message-id: <199603271513.JAA12953@firewheel.cs.utexas.edu>
Subject: Re: Converses (Re: Availability of the ANSI standard proposal?)
To: phayes@uiuc.edu (Pat Hayes)
Date: Wed, 27 Mar 1996 09:13:06 -0600 (CST)
Cc: cg@cs.umn.edu, brayman@zuben.ca.boeing.com, goolsbey@cyc.com,
interlingua@isi.edu, murray@cyc.com, srkb@cs.umbc.edu
In-reply-to: <v02140b04ad7e816e7a9a@[128.174.210.75]> from "Pat Hayes" at Mar 27, 96 00:47:13 am
X-Mailer: ELM [version 2.4 PL25]
MIME-Version: 1.0
Content-Type: text/plain; charset=US-ASCII
Content-Transfer-Encoding: 7bit
Sender: owner-srkb@cs.umbc.edu
Precedence: bulk
>> Bill Brayman writes:
>> In the database world, this issue concerns what are called "views". It
>> is assumed that data is stored in some cannonical form and materialized
>> in various ways depending on it use.
>>
>> So what ontological engineers/scientists are trying to do is reverse
>> engineer real life views to find the ultimate canonical form that will
>> support all views.
>>
>> What Fritz only hinted at is the ontological work that still has to be
>> done to represent views. It vaguely reminds me of Robin Lakoff's (or
>> someone's) observation that kids learn about dogs and cats first and only
>> much later about taxonomies and felines and other more canonical forms.
>> So, what cyc and other systems have to do is be able to make statements
>> about dogs and cats even tho they may be built upon notions of taxonomies
>> and felines and canines.
> Pat Hayes writes:
> I think what Bill is referring to here is that kids learn the concepts in
> the 'middle' of the heirarchies first, dogs and cats before mammals or
> before pekinese. This actually is suggestive that these middle classifiers
> are the most informative: they cut up the set of examples more efficiently
> than the finer ones because they are more evenly balanced. But the more
> things you know about, the higher the 'most useful' splits get to be, which
> is why we tend to make abstractions as we get more knowledgeable.
Hmmm....generalization and canonicalization seem like slightly different
processes to me -- "canonicalizing" doesn't necessarily involve abstraction.
The database analogy is nice -- certainly, database integration involves
trying to merge lots of view-specific data representations into some useful
canonical union from which the views can be regenerated. I think KB
construction involves a similar process too -- but also, and more importantly,
one of generalization, so that domain-specific rules are recognized as
instances of more general principles (eg. work on NEOMYCIN, XCON-IN-RIME).
Presumably the issue of generalizing information doesn't come up in
the DB world much, as all things are instances (but what about OODBs?).
>> In the database world, this issue concerns what are called "views". It
>> is assumed that data is stored in some cannonical form and materialized
>> in various ways depending on it use.
>>
>> So what ontological engineers/scientists are trying to do is reverse
>> engineer real life views to find the ultimate canonical form that will
>> support all views.
I think many people in AI would agree that there isn't some "ultimate
representation" from which task-specific views can be generated -- instead
there are alternative representations/models which can be used depending
on the task. Now presumably this must come up in database integration too:
Sometimes two different views, from two independent databases, may
simply be incompatible. For example: in the oil industry, company A
considers the probability of finding oil as the combination of 4 separate
factors, whereas company B considers it the combination of 6 (different)
factors. These two databases (each containing factors & overall probs)
can't be canonicalized together -- the two companies have simply used two
different ontologies to talk about prospective oil sites. So presumably
in the database world, too, this issue of maintaining multiple and
possibly incompatible representations comes up and has to be addressed.
Best wishes,
Pete
PS:
> Pat Hayes writes:
> This raises another issue, however. Its all very well to say that only one
> ordering needs to be kept and not the converses, but how is it specified
> what the canonical ordering IS? If I write between(d,e,f), what order are
> those points supposed to be in on a line? Where is that information
> represented?
Isn't this simply given by an axiom for between(), eg.
(d < e) & (e < f) -> between(d,e,f)
Am I missing something here?
---
Peter Clark (pclark@cs.utexas.edu) Department of Computer Science
tel: (512) 471-9565, fax: (512) 471-8885 Univ Texas at Austin, TX 78712, USA
Research homepage: http://www.cs.utexas.edu/users/mfkb/dce.html
Personal homepage: http://www.cs.utexas.edu/users/pclark