Re: A simplistic definition of "ontology"
phayes@ai.uiuc.edu (Pat Hayes)
Message-id: <199510062015.AA04617@a.cs.uiuc.edu>
X-Sender: phayes@tubman.cs.uiuc.edu
Mime-Version: 1.0
Content-Type: text/plain; charset="us-ascii"
Date: Fri, 6 Oct 1995 15:30:39 -0600
To: hovy@isi.edu (Eduard Hovy), phayes@ai.uiuc.edu (Pat Hayes)
From: phayes@ai.uiuc.edu (Pat Hayes)
Subject: Re: A simplistic definition of "ontology"
Cc: srkb@cs.umbc.edu, cg@cs.umn.edu
Sender: owner-srkb@cs.umbc.edu
Precedence: bulk
At 12:04 PM 10/5/95 -0500, Eduard Hovy wrote:
>At 2:15 PM 10/5/95, Pat Hayes wrote:
>>At 2:59 PM 10/4/95 -0500, Eduard Hovy wrote:
>>>What does "specify a concept" mean, if not just to list the relationships
>>>of the concept with other concepts?
>>
>>It means giving a structured *theory* of those relationships, ......
>Words like "theory" (and especially "*theory*" :-) ) confuse me;...
Let me try clarify terminology and then to make my original point more
clearly. Technically, in logic, a 'theory' is a vocabulary - ie a set of
names for things and names for relations - and a finite set of axioms
written using that vocabulary. (If its allowed to be infinite we move
beyond first-order logic, but this is often a fairly small step if the
infinite sets can be somehow finitely characterised, eg by an axiom
schema.) The theory captures all the information that is there: one can't
appeal to any 'external' meanings that the vocabulary might have. As others
have noted, it is often a good exercise to replace all the names by gensyms
and then see how much sense it makes, because thats all the sense it makes
to any mechanical processor.
(In KR applications the 'theory' of a set of concepts will often not fit
this definition exactly, since the axioms might use vocabulary from other
theories. Perhaps we should speak of a theory 'relative to' an external
vocbulary, or some such locution. We can usually be sloppy about this, but
there are times when it is important. In any case, the 'entire' KB will
still be a theory, at least until we consider inputs from sensors and so
forth. For more discussion on this, see my 'second NP manifesto'.)
In the modern spirit of representational ecumenicalism, every 'theory'
doesnt have to be a set of axioms - the 'knowledge' might be encoded in all
kinds of ways - but whatever it is, it ought to be able to do what this
kind of theory does, more or less, which is to produce (or somehow act as a
validation for) all the consequences of the information it supposedly
represents.
Now, when the modern Ontology church was started a few years ago, it was
clear that for many people, including our founding father, Ontologies
differed from AI knowledge bases (such as CYC) precisely in that they could
contain information written in ordinary text, information which definitely
was not part of any such 'theory'; and this information was to be
considered not just a kind of helpful commentary for the human reader, but
an integral part of the Ontology, sometimes the main part. By a 'glossary'
I meant an Ontology which is completely devoid of any theory (in the above
sense) and consists entirely of such 'comments': a kind of organised,
heirarchical dictionary which is explicitly intended for use by humans, and
which makes no claims to 'represent knowledge' for any AI purpose. The
original call for papers was careful to insist that this was a genuine
Ontology,and no 'axioms' or anything like them were required.
I think that there are two rather different perspectives. For one, the
English (or Dutch, or whatever) text is a promissory note, as it were (as
you say, perhaps its the best we can do at present), but the ultimate goal
is to somehow get all of its content down into the theory, accessible to
mechanical processing. This is the 'KR' side of the gulf. From the other
persective, Ontologies have nothing particular to do with what AI and NLP
call knowledge representation: they are ways of making human communication
more precise by imposing structure on parts of natural language, especially
specialised technical vocabularies. From this perspective, formal theories
such as axioms are thought of as maybe giving a kind of especially sharp
focus on the terminological interactions, but they never taken to
completely embody the meanings of anything. To replace the terms in such a
glossary by gensyms would be crazy, because humans would then be able to
read it. Issues of 'readability' of logical syntax keep emerging in the
discussions, for example; on this side, readability is crucial, while on
the KR side it is of little importance; at most only a matter of syntactic
sugar. (KIF explicitly renounces any claim to be readable, but is regularly
criticised for being unreadable.)
My worry is that maybe we have here two rather different kinds of thing
here, and criteria of success, etc., are different in the two cases; and if
so, we ought to get our ideas clearer.
I would be interested if anyone thinks that this distinction is somehow
mistaken. Does anyone else feel this tension between the two perspectives?
Pat
------------------------------------------------------------------------------
Beckman Institute (217)244 1616 office
405 North Mathews Avenue (415)855 9043 or (217)328 3947 home
Urbana, Il. 61801 (217)244 8371 fax
Phayes@cs.uiuc.edu