Top level ontology

John F. Sowa (
Sun, 30 Nov 1997 11:45:51 -0500


I'm glad that people are finally turning their attention to the top levels.
Many people, including Doug Lenat and others, have argued that the top levels
aren't important because the significant inferences are generated by the
intermediate levels.

I will agree that that conclusion is probably true for many systems, including
Cyc, largely because they don't have a well-designed top level. I agree with
the Cyc principle that all of the categories in the ontology should have
axioms with significant implications. Categories should not be tossed into
the hierarchy just for decoration, although I will continue to insist that
a well-designed hierarchy should have a significant amount of symmetry.
But that symmetry isn't the original motivation for the choice of categories;
instead, it is the result of asking some basic questions:

1. Whenever there are two or more subtypes of any given type, there must
be some reason for making that distinction. What is it?

2. Given any distinction made at one part of the hierarchy, is that
distinction unique to the branch where it happens to occur? If so,
what makes it unique? If not, are there any other branches to which
it might apply? If so, has it been applied there? If not, why not?
If so, can the distinction be generalized to a higher level?

3. Any distinction that is used to create branches is useful for
classification. But does it also have implications for inference?
It would be extremely odd if there were features in the world, in the
biological kingdoms, or in language that had no implications whatever.
One of my favorite quotations was made by the linguist Dwight Bollinger:
"Every difference makes a difference." If you ever discover any
distinction that is used for classification, that distinction almost
certainly has implications. And those implications determine the
associated axioms.

There is much more that could be said about this topic, but the basic point
is that an unsymmetric hierarchy suggests that there are missing distinctions.
Sometimes a missing distinction might not cause obvious trouble, since the
only effect would be to eliminate conclusions that could otherwise be drawn.
But often, a missing distinction could create inconsistencies because it
could lead to misplaced branches in the hierarchy, which would cause
spurious inferences.

There is also a very practical reason for moving general distinctions as high
up in the hierarchy as possible: if you observe that all physical things
have positive mass or energy, then you don't have to repeat that axiom.
It also gives you practical guidelines for knowledge acquisition:
if a category C has N supertypes, each of those supertypes has implications,
which can be used as tests for whether a particular entity should be
classified under C. Knowledge acquisition tools can automatically generate
questions from those guidelines to prompt the user or knowledge engineer.

There are also some issues that you and I have never completely resolved.
One of them is the principle of identity, which I consider a matter of
convention more than a matter of how the world happens to be constructed.
I don't consider these distinctions to be distinctions about how the world
is, but distinctions about the language or semiotic system that we happen
to be using for describing the world. All my categories are signs that
are used for language, natural or artificial, and I don't believe that
we can ever or should ever develop language-independent categories.

As you know, I consider Peirce to be one of the top five ontologists
(the others being Aristotle, Leibniz, Kant, and Whitehead). And I don't
believe that any serious ontology can be constructed without taking
Peirce's semiotics as the foundation.

Recently, I've been reading the book _Charles Peirce's Pragmatic Pluralism_
by Sandra Rosenthal (State University of New York Press, Albany, 1994,
available in paperback for $16.95). It is especially good for its references
to Peirce's original writings and to other commentators who have commented
on his work, rightly or wrongly. Following is a passage that applies
to our discussion (page 13):

Focussing on a passage from Peirce in which he stresses "How much more
the word _electricity_ means now than it did in the days of Franklin;
how much more the term _planet_ means now than it did in the time of
Hipparchus," Hjalmar Wennerberg objects that Peirce's theory "blurs
the important distinction between logical analysis and empirical research."
To this, Skagestad responds that "Peirce does not blur this distinction
in the least; he unconditionally denies that there is any such

Rosenthal goes on to clarify Peirce's position as being very close to
Kuhn's view of scientific revolutions, but with some important improvements
on Kuhn. One very important implication for our work on ontology is that
you cannot do work on theory and empirical studies separately. In the
ontology meetings, I have seriously objected to drawing a distinction between
a "theory group" and an "empirical group". I think that kind of split is
extremely harmful, both organizationally and practically. You can't even
have meetings that are devoted to one topic or the other: every discussion
of theory must relate to empirical data, and every discussion of the data
must involve theory. The same people must do both.

John Sowa