Re: Good and Bad IS-A hierarchiesPeter Clark <firstname.lastname@example.org>
From: Peter Clark <email@example.com>
Subject: Re: Good and Bad IS-A hierarchies
Date: Thu, 8 Jun 1995 14:10:53 -0500 (CDT)
In-reply-to: <firstname.lastname@example.org> from "Eduard Hovy" at Jun 7, 95 04:45:19 pm
X-Mailer: ELM [version 2.4 PL24alpha3]
Content-Type: text/plain; charset=US-ASCII
>>> Peter Clark wrote, among other things, this:
>>> What I wouldn't like to now see is a lot of isa-hierarchies published,
>>> with little or no information in the general/top-level concepts.
>> Fritz's comment:
>> The key thing is the "little or no information". I _would_ like
>> to see generally useful IS-A hierarchies published, provided they
>> explicity state what information is contained in each concept.
> Ed Hovy wrote:
> So I disagree with Peter Clark. I *would* like to see a lot of concept
> organizations (isa hierarchies at the very simplest), presented in a way
> that is easily browsable so that I can pick and choose bits and pieces
> from here and there and weld them together to suit my needs of the moment.
I hope I didn't sound like I was against concept organizations -- concept
organizations (or rather representations) are what it's all about, I'd like
to see more. Fritz's interpretation is right, my concern was writing down
concept names without providing any information about them. Just because an
ontology has a frame named "container" (say) the author shouldn't automatically
conclude that he/she's represented the notion of containers -- what McDermott
calls "wishful mnemonics" ("AI meets Natural Stupidity" in "Mind Design",
Ed: J. Haugeland, 1981, MIT Press 143-160). The representation of "containers"
should (ultimately) be richer than a token in a hierarchy -- for example
it might include statements like containers bound a space, that things
can be in or out of it, that things can be put in / got out, that you can't
get something out unless it's already in, that they might have portals which
can be open or closed, etc. (Pat Hayes did something along these lines in
"Ontology For Liquids" in "Readings in Qualitative Reasoning", as has Cyc
Penman's ontology is an interesting case -- I'll comment on it although
at the risk of mistakenly describing it as I'm not an expert on it.
It looks on first glance like an isa hierarchy with little information
in the top level concepts. How then can it be used for sentence generation
if there's no representation of what the terms in the upper model mean?
The answer is that the meaning is implicit in the Penman program itself
which uses the ontology to generate sentences. It's the program
which defines what it means -- in sentence generation terms -- to be
an "object", "process", or "quality" (say) (which could just as well be
called "type1-words", "type2-words", "type3-words"). It'd be nice if
these meaning implicit in the Penman sentence generator could be made
accessible / explicit / exportable to other systems. That's really the
"knowledge" which would be nice to share.
So, okay, I agree isa hierarchies are the simplest way to start --
but provided they're recognized as exactly that, namely a start rather
than a complete representation. To quote Ben Kuipers: "taxonomy doesn't
Peter Clark (email@example.com) Department of Computer Science
tel: (512) 471-9565 University of Texas at Austin
fax: (512) 471-8885 Austin, Texas, 78712, USA.
Project homepage: http://www.cs.utexas.edu/users/mfkb