Re: clarifying clarifying ontologies

"Kenneth D. Forbus" <>
Date: Mon, 7 Aug 95 14:51:25 CDT
Message-id: <>
X-Mailer: Windows Eudora Version 2.1.1
Mime-Version: 1.0
Content-Type: text/plain; charset="us-ascii"
To: (Pat Hayes), (Eduard Hovy), (Pat Hayes),
From: "Kenneth D. Forbus" <>
Subject: Re: clarifying clarifying ontologies
Cc: (Fritz Lehmann),,,
Precedence: bulk
With some trepidation, here's my $0.02.  As far as I'm concerned, Pat and
Doug are right.  Taxonomies of "concepts" (or even "predicates") without
axioms (or their moral equivalent) that pin down their intended meaning are
not at all useful.  It makes one feel good to write them down, for some
reason, but it doesn't lead to anything.  Take a look at KR papers from the
1970s.  Many of them had this flavor.  Then when it came time to start
really laying the knowledge out, i.e., the hard part, well, alot of people
bailed, alot became logic-hackers, and so on.  

If someone hadn't mentioned compositional modeling earlier, I doubt that I
would have chimed in.  One thing you'll notice about compositional modeling
is that it doesn't rely on taxonomies.  In fact, while sometimes I've gone
back and drawn a taxonomy after the fact to make things clearer for a paper,
I've never found taxonomy-creation to be a critical activity in building a
successful domain theory.  Axiomatizing predicates, on the other hand, is
where the real work is.
Our current "hard-core" quantitative thermodynamics KB has maybe a few
hundred axiom-equivalents in it, which suffices to analyze a huge number of
thermodynamic cycles, but I have never figured out what the taxonomy of
concepts in it is.  Despite that, I believe the KB is extremely clean and
well-modularized.  The evidence for that is that Peter Whalley (my
collaborator, who is a thermo hacker) and I could add the knowledge needed
to analyze non-steady flow cycles in only two days, with minimal changes to
the existing knowledge.

Pat's idea of clusters is, to me, a much better way to proceed.  It is
certainly what my group has been doing for about a decade now, with
reasonable success.
If you look at our work in mechanics, (H. Kim's thesis for instance), you'll
see that there is a growing body of qualitative concepts of surface, vector,
etc. that can be used for all sorts of things.  (Latest application attempt
I've heard of is modeling erosion of river banks on South African game
The reason this knowledge is useful isn't because we've agreed on names for
predicates, but because we've laid out a set of axioms/rules that describe
what inferences they sanction, and demonstrated by programs that use them
that these inferences are enough to do some useful things.  That's where the
payoff is.

I don't want to be too negative.  I think the growth of a community that
takes knowledge representation seriously, that is, is actually committed to
REPRESENTING KNOWLEDGE in ways that can be combined usefully to ultimately
create the kind of understanding of intelligence and very smart software
that we all ultimately want, is a wonderful thing.  But I ask that you
please don't postpone diving into the deep waters of actually fleshing out
the knowledge in some particular area(s) in favor of spending all your time
splashing around in the shallows of taxonomy creation.  It takes good lungs
to spend alot of time underwater, but with practice it comes easier, and in
fact it's quite wonderful down there!

Prof. Kenneth D. Forbus
Qualitative Reasoning Group
The Institute for the Learning Sciences
Northwestern University
1890 Maple Avenue
Evanston, Illinois, 60201, USA

voice: (708) 491-7699
fax: (708) 491-5258