merging ontologies: a thread
Tom Gruber <gruber@HPP.Stanford.EDU>
Message-id: <199405251002.DAA18379@HPP.Stanford.EDU>
X-Sender: gruber@ksl.stanford.edu (Unverified)
Mime-Version: 1.0
Content-Type: text/plain; charset="us-ascii"
Date: Wed, 25 May 1994 03:01:57 -0800
To: srkb@cs.umbc.edu
From: Tom Gruber <gruber@HPP.Stanford.EDU>
Subject: merging ontologies: a thread
Sender: srkb-owner@cs.umbc.edu
Precedence: bulk
One of the central issues for knowledge sharing is how to merge or combine
existing ontologies. Here's a recent thread on the subject in which the
subject of semiformality also got dragged in. Posted with permission.
From: vet@cs.utwente.nl (Paul van der Vet)
To: fritz@rodin.wustl.edu, gruber@HPP.Stanford.EDU
Date: Wed, 27 Apr 94 15:25:44 +0200
Subject: Combining ontologies
In comp.ai, you posted the following question:
"At present there are no automated methods
to combine ontologies, no formal theories
for combining ontologies, and no tools available
for combining ontologies except for hierarchy-
browsing assistants."
Does anyone have counterexamples to this answer?
I don't, and I would be very surprised if someone had. Perhaps I'm
mistaken, but I see no flaws in the following argument. Please
correct me if I'm wrong.
In Gruber's enlightening explanation of what an ontology is (KR'92, if
I remember correctly, and an internal KSL report), the definitions
making up an ontology are comprised of a formal and an informal part,
and both are indispensable. You design a formal structure but will
always need primitives whose meaning is explained by anchoring them
into domain knowledge (for example). This can only be done by a
natural language account. Demanding that those meanings, too, are
pinned down by formal definitions would only lead to a regress,
because for those definitions you would need new primitives. And so
on.
Now an automated system able to combine ontologies has to determine
whether the concepts of the two ontologies can be mapped onto each
other. To be able to do that, the system has to reason about the
meanings of the primitives. One way to do that, of course, is by using
a pre-determined third ontology covering both - but that wasn't the
point, was it.
From: Tom Gruber <gruber@HPP.Stanford.EDU>
Date: Sun, 15 May 1994 12:59:25 -0800
Subject: Re: Combining ontologies
I agree that no formal machinery will be able to completely automate the
combination of ontologies that contain primitive (incomplete or informal)
definitions. I take "combination" to be something like merging the
definitions with consistent renaming and the construction of axioms that
map terms in one ontology to terms in the other. Formal machinery can
help, but will be incomplete when there are primitives.
This doesn't mean there can be no tools to help. Just as there are tools
that help humans do natural language translation, there can be tools that
help humans translate and merge ontologies.
From: vet@cs.utwente.nl (Paul van der Vet)
Date: Mon, 16 May 94 16:12:51 +0200
Subject: Re: Combining ontologies
> Formal machinery can help, but will be
> incomplete when there are primitives.
Just to be sure: are you aware of ontologies that have no primitives
in their formal part?
From: Tom Gruber <gruber@ksl.stanford.edu>
Date: May 23, 1994
No, I'm not aware of any. The KIF ontologies start out with primitive
concepts of set and list, and build up machinery from there. If anyone
ever wrote down an axiomatization of numbers base on set theory, it could
be free of primitives. However, in all practical settings of knowledge
sharing that I can imagine, humans will have to understand the intended
meaning of the concepts using some knowledge beyond the constraints
explicitly there in the axioms. (A fancier way to say this is that the set
of intended models of the theory is almost always a subset of the possible
models in the standard Tarksian semantics.)
I take the point of view that the definitions in an ontology are there for
communication, since the purpose of the ontology is to specify a shared
conceptualization. For the purpose of effective communication, axioms
ain't all there is. Intended meaning can be explained in other media, such
as natural language and working programs ("operational semantics").
Nonetheles, axioms are great if you can get them. There is a reasonable
argument that natural language is superior for the task of writing
ontologies (Skuce, <199203031953.AA01955@quark.isi.edu>), since we often
don't know what we're talking about and natural language lets us get away
with it. But taking that path I could argue that demo programs with cute
graphics also convey meaning, and should be the preferred medium for
ontologies ("demonstrational semantics"). The case for axiomatic accounts
is both theoretical and practical. The theoretical claim is familiar: the
meanings of logical sentences are context-independent (e.g., independent of
culture, "situation", time, and computer implementation). The practical
advantage is that we can do analysis on them (in many cases, with software
assistance).
So, if you buy the last two paragraphs --- that axioms are both
insufficient and the best we've got --- then I'd offer that a hybrid of
axiomatization and clear writing (and demonstrating) are a good way to
write ontologies. Admittedly, this is a strategic position and not a
scientific claim. If anyone knows of criteria on which we can measure the
effectiveness of various media for communicating intended meaning, we can
put it to a test.
tom