Re: Web Rule Language - WRL vs SWRL

> [Bijan Parsia]
> The old Layering Story has been bankrupted in several different 
> ways.

I'll take advantage of this crack in the orthodox Wall of Opprobrium
here to jump to Michael Kifer's defense.  Someone should.

For one thing, I find the argument against his position to sound a
little like this one:

"Don't mix coal dust with the pixie dust --- it burns smoky and
pollutes the air."

Yes, pixie dust is much better than coal, except that it doesn't
exist.  I believe that there are applications in which the logical
purity of Owl, DLP, etc. can be preserved, but (as Michael has
implied, I think) there are many domains for which people want rules
in order to express problem-solving strategies, and I have serious
doubts that those will be expressible in pure FOL.  The reason why
nonmonotonicity is (regrettable or not) ubiquitous in real-world
applications is that most inference is not deductive.  Any attempt to
reduce an inference task to deduction is either completely misguided,
or ends up compromising the purity of the deductive notation.  Feel
free to outlaw this practice; you won't succeed.  If the semantic web
succeeds, one reason will be that someone implements a "killer app"
rule system that solves a lot of realistic problems, and it will
surely have a quasi-procedural --- and nonmonotonic ---
interpretation.

Speaking of semantics, there is something about Ian Horrocks's
oft-repeated claim (to pick one recent iteration) --

  "An LP language would find an entailment that is *not* supported by 
   RDF semantics .... Ergo, LP is semantically incompatible with RDF."

-- that I find dubious, in spite of its apparent obviousness.  Suppose
one adopts a standard Tarskian account of semantics.  It seems to me
that a monotonic and nonmonotonic system could have exactly the same
Tarskian semantics, and still disagree on what followed from a given
set of facts (call it S).  That's because it's only in classical logic
that the facts entailed by S are exactly the facts that follow from S.
In a nonmonotonic system the set of facts that are inferred from S is
a superset of the set of facts that are entailed.

This point may be obscured by the way LP is presented nowadays.  In
the past two decades logic programming and logic have drifted apart,
to the point where their are routinely characterized in basically
different ways.  I think this is unfortunate.  I prefer to think of
nonmonotonic systems as being "basically" ordinary logic, with an
escape hatch here or there.  I grant that this is hard to formalize,
but that's because it's precisely the nondeductive parts of inference
that are hard to find a uniform formal framework for.  But consider
this hypothetical example: a logical system with procedural hooks to
connect to a Bayesian inference system.  It might have a nonmonotonic
rule to the effect that "If the best estimate of the probability of P
is > 0.9, infer P."  One could develop a pretty formal account of how
this thing worked, including an account of what inferences were
licensed under what circumstances.  But it would have nothing to do
with the _semantics_, which could be specified in advance,
independently of the details of the inferential mechanisms.

                                             -- Drew


-- 

                                         -- Drew McDermott
                                            Yale University
                                            Computer Science Department

Received on Thursday, 30 June 2005 02:29:51 UTC