Re: rdf inclusion

   [Jeff Heflin]
   My personal opinion is that if you're using an ontology language, every
   term you use must be defined in some ontology (even if only to say that
   it is a class or property). 

I agree with the sentiment, but please let's not use the word
"definition" the way you and Dan B. are using it.  Ontologies express
relationships among terms, but they almost never define them.  People
colloquially speak as though a statement like "living things are
partitioned into vegetables and animals" is a definition of,
say, "vegetable" and "animal," because it is, in some sense, a
"declaration" of these symbols, and in the computer world it is
usually the case that declaring something is necessary and sufficient
to define it.  But (as I know you know) in a KR language that is not
the case.  In fact, as Pat Hayes has argued, it is hard to say what the
*logical* difference is between the "paritition" statement above and
the seemingly humbler statement that "Sally is a vegetable."

But you're completely correct that importing an ontology is different
from pointing to a web page or even a set of assertions.  At the risk
of repeating what you said, here are the two key reasons why:

1. The purpose of an ontology is to allow agents to draw conclusions,
   and in particular to detect inconsistencies (e.g., type errors) in
   datasets.  A dataset without its associated ontology has no
   rationale that I can see.  If I give you a set of facts about
   numbers, but I view Peano's axioms as an option that you can take
   or leave, then the set of facts could be taken to be saying
   anything at all about any topic at all; what in the world is the
   point? 

2. Ontologies are *small* and *internally consistent.*  I think the
   argument about importing assumes that once I start following
   pointers to the web I could wind up anywhere.  That may be true in
   general, but it's not true for ontologies.  Whoever designed an
   imported ontology didn't just throw together some stuff they found
   on Google.  They had to think through many tough questions, and at
   every stage slight changes from the design decisions taken would
   have introduced subtle gaps or inconsistencies.  Chances are the
   designers had to backtrack several times as these pitfalls were
   encountered.

Point (2) has as a consequence that if my dataset imports Ont-1, I can
be confident that all the ontologies Ont-1 imports are, in the
designers' minds, *coherent pieces of Ont-1.*  If they weren't, the
designers would have imported something else or built what they
needed.  Furthermore, the chain of imports is likely to be shallow;
and if Ont-1 imports Ont-2 and Ont-3, following the import links from
Ont-2 and Ont-3 is likely to get you to a common ancestor more
frequently than chance would predict.  Coherent theories just don't
look like balls of string.

                                             -- Drew McDermott

Received on Thursday, 25 April 2002 15:36:33 UTC