Re: rdf inclusion

"R.V.Guha" wrote:
> 
> Jeff Heflin wrote:
> 
> >I believe that it was intended for such an environment, but IMHO it
> >falls well short of those goals. In order to handle a distributed
> >environment, you need more than just "all identifiers will be URIs." You
> >need some way to come up with shared definitions even though global
> >agreement is impossible. To make this machine processable, you need a
> >formal semantics that explicitly takes into account how the distributed
> >axioms used to reason about a set of documents is assembled. Otherwise,
> >different reasoners can conclude wildly different things from the exact
> >same information, providing us absolutely no guarantees about
> >interoperibility. Now I understand that RDF was designed to be a
> >foundation which a lot of things could be eventually built on (I've seen
> >the "layer cake" diagram a million times). However, I think that the
> >distributed nature of the Semantic Web should be a fundamental issue,
> >and not just an "add-on." Furthermore, now that people are beginning to
> >build some of these upper layers, they are finding that RDF is not as
> >easy to extend as they might have wished: the WebOnt WG is already
> >discovering that extending RDF with additional semantics is not a
> >trivial matter.
> >
> >
> Jeff,
> 
>  Different reasoners *will* reach different conclusions, depending on
> what their inputs are. This is
> just the way the real world is. People (and hence reasoners) will have
> different theories of who shot
> JFK, whether Bin Laden is still alive, ... If  we architect something so
> rigid that the entire system had
> to agree on one answer for each of these questions, I doubt it would be
> very interesting.

Of course different reasoners will reach different conclusions if their
inputs differ. What I'm worried about is that different reasoners might
reach different conclusions even when given identical inputs. I
certainly wouldn't suggest that there is one set of answers for every
question on the Web. However, I think it is absolutely critical that if
I give two sound and complete reasoners the same set of documents, that
they agree on all answers to all queries about them. You can't blindly
accept everthing, but if we allow these reasoners to arbitrarily pick
and choose what they believe in these documents, then we may as well not
even have semantics in the first place, because there will be no way to
predict what the answers to the systems will be. I suggest that in
addition to traditional model theory, semantics for web-based systems
must have explicit rules for what you can infer when you combine
documents. I believe Pat Hayes is suggesting a similar thing elsewhere
in this thread when he talks about "	public rules of inference
behaviour." Having things like "daml:imports" that explictly link
documents or ontologies help define these rules. 

>  You also make the statement that "formal semantics" (by which I assume
> you mean some kind of
> tarskian model theory) will provide the answer to the problems of
> integrating data from different
> sources. This is a very very strong claim that is being made explicitly
> and implictly, that is as yet
> unjustified. I would really like to understand how a model theory will
> solve these problems.

I think your work on context logic has some ideas that are very relevant
to the Semantic Web, and is one of things that influenced my thesis [1].
If you have the appropriate "lifting axioms" (to borrow a term from your
thesis), then you can integrate heterogeneous data formats.  Now of
course this approach is limited by the expressivity of the language in
which you can express your axioms in, but it is a good start. Could you
suggest an alternative approach that is not grounded in some formal
semantics?

>  About the issue of RDF & RDFS being hard to extend --- let us be *very*
> clear on this. RDF & RDFS
> were designed to be Cyc like systems [1]. They were *not* designed to be
> DL like systems. You are finding
> it hard to reconcile the two. Cyc-like systems are extensible and have
> been extended, though not in a
> fashion that is consistent with DL model-theories. Yes, the clothes
> don't fit the person. Maybe the problem
> is with the clothes and not the person.

Don't get me wrong, I'm no huge fans of DLs as the basis for the
Semantic Web (although I've gradually become aware of certain benefits
of that approach). My original work in the area (SHOE [2]) was based on
a datalog-like language. If I had my choice, WebOnt would be starting
from a similar basis. However, the fact is that RDF(S) is not even easy
to extend correctly for a simple rule-based language. In order to write
an implication or a disjunction and not assert all the atoms you have to
resort to reification or look at arcane things like "dark triples" which
is currently being considered by the WebOnt WG. Since Cyc is much more
expressive than rules languages, I highly doubt that you can write an
extension of RDF that preserves RDF semantics while adding the semantics
of Cyc without going through many of the same troubles as the WebOnt WG.
However, I'd be perfectly happy to be proven wrong.

Jeff

[1] http://www.cse.lehigh.edu/~heflin/pubs/#heflin-thesis
[2] http://www.cs.umd.edu/projects/plus/SHOE/

Received on Thursday, 23 May 2002 14:36:25 UTC