Re: [SUO]: Key questions about common upper ontologies
John Sowa makes several points about whether and
how users can adopt a common upper ontology.
I think there is more to be said about the
issues:
(1) Can "we" "legislate" an upper ontology?
(why not say "recommend"??)
(2) How come WordNet is more widely used than CYC?
(3) Do different purposes actually require logically
inconsistent representations?
-- therefore will say it:
John F. Sowa wrote:
**** > (1)
> Pat and Danny,
>
> I sympathize with Pat's hopes, but I agree with
> Danny that we are not going to get there by a
> brute-force legislation of one universal ontology.
>
"we" (the IEEE-SUO group) won't get there without
serious funding (which has never been available to
this group or any other broad-based group trying
to create a standard upper ontology), but
Jim Schoening's note asked about whether it would
be feasible within one organization of many branches.
In that case it is feasible, and I would highly
recommend it.
**** > (2)
John also used CYC vs. WordNet as an example of
what might be adopted.
> There are two case studies that indicate where
> the "sweet spot" or the "low-hanging fruit" is
> available, and Danny is closer to it than Pat:
>
> . . . .
> Since OpenCyc has been released, many groups have
> started to use it, but they are using it more like
> WordNet than like the fully axiomatized Cyc. That
> is partly because OpenCyc does not have all the
> axioms even for those terms that have been released.
> However, I am also aware of some research projects
> that have access to the full Cyc system, and many of
> them are only using the hierarchy in a WordNet-like
> approach, rather than the fully inferential way that
> the Cyc axioms are designed to support.
>
Fine, but there are three perfectly good reasons why
WordNet might be used at this time more than CYC:
(1) CYC is proprietary, all of WordNet is free.
Even the fact that the openCyc is free to use
but still copyrighted may leave some potential
users hesitant to get into ambiguous legal waters.
More importantly, the openCyc, as John mentions,
has few of the axioms that make CYC logically powerful.
But it does have all of the terminological obscurity that
would repel almost any potential user that is not
absolutely determined to build a serious ontology for
serious reasoning purposes.
(2) CYC is harder to understand than WordNet.
How many people outside of CYCorp know what
#$relationAllExists means, and why it is very
important?
It is therefore harder to use and therefore:
(3) The fact -- (I accept John's assertion as a fact,
it sounds reasonable) that most users of CYC
use only the hierarchy -- merely shows that it is
easier to develop a computer system that does
very shallow reasoning than one that uses the
kind of deeper inferencing that CYC was designed
to enable. I am not surprised or dismayed that
simple programs are much more common than
complex programs.
It is also much more likely, for someone
adopting an ontology for an application, that
one will find terms closely related to one's
domain of interest in WordNet (> 100,000 words)
rather than OpenCyc (3,000 classes, 2,000
relations). For shallow reasoning like counting
links, it has the built-in hierarchy and enough
other relations to keep programmers busy for years.
But if one wants to do deeper inferencing, CYC
has some very nice features that are absent in WordNet.
Stay tuned. In developing cognitive systems, we
are still in the stage where Goddard was in developing
liquid-fueled rockets -- they could reach a mile in height
and were developed by a team of a few people. We haven't
gotten to the V-2 stage and are still quite a distance from
a moon shot. More money, please. V-2 was only developed
due to the pressure of war, and the Moon Rocket due to
the cold war. We don't have any driving pressures
of that kind now -- unless we can convince the legislators
that the kind of communications failure of our
intelligence agencies prior to 9-11 could be ameliorated
by a common upper ontology. I would be willing to make that
claim.
An upper ontology that has any chance of being
widely adopted for complex reasoning purposes will have
to be at least as complex as the openCYC. I myself would
use a merger of CYC and SUMO, plus some concepts from other
ontologies, such as those included in the MSO proposal.
**** > (3)
> PC> We need not debate whether it is **possible** to
> > represent the same thing in logically inconsistent ways
> > -- it does seem so, if one never encounters a situation
> > where the differences have practical consequences -- but
> > is there actually a **need** to do so? For what purpose?
>
{JS]
> Nobody (except Microsoft, perhaps) deliberately tries to be
> inconsistent with anybody else, but what is important for
> one application is often irrelevant for another. For example,
> we might say that a ball is spherical. But no actual ball
> is really spherical, and a soccer ball, baseball, tennis ball,
> or golf ball has very clear and well defined indentations,
> ridges, and roughness that deviate from a sphere. And a
> football or rugby ball are not even designed to be spherical.
> A ball bearing is one of the few things for which a perfect
> sphere is ideal, but none of them reach that goal. And
> nobody even attempts to make a perfectly spherical meatball.
>
Any ontology that cannot represent approximations will be
quite hopeless for reasoning about the real world.
Users can chose to adopt whatever level of precision is required for
their application. It is important not to mix in the
vague usage of language with the logical specifications of an
ontology. Mapping from language to the ontology is important
for applications that handle language, but separate from the question
of whether the ontology can represent all of the different concepts
that may be labeled by a single word.
CYC does have some provisions to handle approximation issues. Consider
just one of the OpenCyc relations:
(#$comment #$canContainShapes "The predicate #$canContainShapes is used to give
an approximation of the internal size and shape of particular tangible objects,
by relating an object to an abstract region of space described as a geometric
shape with definite dimensions. (#$canContainShapes OBJ SHAPE) gives an upper
bound for the size of things that can be contained in the object OBJ, by
specifying the dimensions of an abstract shape which OBJ can contain.
#$canContainShapes uses the elements of #$ShapeFunction (q.v.) for reference,
especially the basic shapes generated by #$RectangularSolidFn, #$CylinderFn, and
#$SphereFn. For example, the trunk of my Honda Civic #$canContainShapes of
(#$RectangularSolidFn (#$Meter 1) (#$Meter 0.5) (#$Meter 0.75)). Cf. #$fitsIn.")
(#$arg1Isa #$canContainShapes #$PartiallyTangible)
(#$arg2Format #$canContainShapes #$SingleEntry)
(#$arg2Isa #$canContainShapes #$GeometricThing-Abstract)
(#$argFormat #$canContainShapes 2 #$SingleEntry)
(#$arity #$canContainShapes 2)
(#$genlPreds #$canContainShapes #$spatiallyRelated)
(#$isa #$canContainShapes #$BinaryPredicate)
(#$isa #$canContainShapes #$StrictlyFunctionalPredicate)
(#$relationAllInstance #$canContainShapes #$TrainCar (#$RectangularSolidFn
(#$Foot-UnitOfMeasure 14) (#$Foot-UnitOfMeasure 14) (#$Foot-UnitOfMeasure 49)))
(#$relationAllInstance #$canContainShapes #$RoomInAConstruction
(#$RectangularSolidFn (#$Meter 4) (#$Meter 3) (#$Meter 2.5)))
(#$strictlyFunctionalInArgs #$canContainShapes 2)
It is logical specifications of this kind as well as
others such as the related relation #$fitsIn (and some I would
add that are not in CYC) that will allow ontology developers
to provide a single ontology that can express
concepts as vaguely or precisely as necessary for particular
applications. Measurements at different levels of precision
are **NOT** logically inconsistent if the notion of the precision
of a measurement is included in the ontology, as it must be.
In a similar vein, different physical theories such
as Newton's laws of gravity and Einstein's general relativity
are **not** logically inconsistent because they are not
intended to be mathematical hypotheses that are of
infinite precision, but are laws expressing empirical
relations that are believed to apply to a certain level
of precision (and which may or may not suggest something
about the ultimate nature of the systems described).
Einstein's equations are more precise for certain
extreme situations, but they may also be ultimately
superseded by a more general theory. Even now it is
rare for anyone to do a calculation of near-earth
trajectories using Einstein's gravity -- Newton works just
fine, even though we know it's not as precise as Einstein.
It's a good-enough approximation that's easier to use.
We can express such better and better approximations in a single
logically consistent ontology.
There will of course be cases where scientific theories are not merely
superseded by better approximations, but proven flat-out wrong.
There was, for example, a theory at one time that it was a protein
that carried the genetic code from generation to generation.
Every ontology must have a provision for alternate theories,
differing beliefs, contexts, etc. and a provision for belief
revision. These will be accommodated in a single logically
consistent ontology provided that assertions are always made relative
to some context. The default context would be the real world
(as CYC puts it "consensus reality"). If assertions about the
real world prove to be logically inconsistent, then we would have
to create alternative theories that are logically inconsistent
**only with respect to those assertions and their
inconsistent consequences**. This is important, because it is only
by trying to maintain a single logically consistent upper ontology that
we will be able to detect and minimize the logical inconsistencies that
may creep in as knowledge is added, and we will have the
ability to isolate and sequester those logical inconsistencies in
minimal contexts that allow consistent reasoning when those
specific inconsistent assertions are not important for our purposes.
[JS]
> I'll end by quoting Danny's conclusion, which I support:
>
> DA> Successful communication does not require the whole abstract
> > language or concrete terminology. This is advantageous to the
> > adoption of any SUO, as people are more likely to be willing
> > to augment their existing languages/vocabularies than replace
> > them with a whole new language infrastructure.
>
No argument here. I discuss this at greater length in my previous note.
Pat
=============================================
Patrick Cassidy
MICRA, Inc. || (908) 561-3416
735 Belvidere Ave. || (908) 668-5252 (if no answer above)
Plainfield, NJ 07062-2054
internet: cassidy@micra.com
=============================================