Proposed response to Martin Merry, HP

The following is my proposed response to the message from Martin 
Merry, which is found at:
http://lists.w3.org/Archives/Public/public-webont-comments/2003May/0046.html


Thanks for your comment.  We believe it raises two important 
questions, one is whether the design of OWL DL as is would require a 
long CR period, the other is whether dropping #hasValue and #oneOf 
should be done by the WG.  Let me answer these in reverse order.

1 - a Long CR:
  From a Process point of view (and ignoring your technical point for 
a moment, I promise to return to it), I would point out that there is 
no requirement that all features of a recommendation must be handled 
by all implementations.  The process document states that the 
requirement to move from CR to PR is:

each feature of the technical report has been implemented. 
Preferably, the Working Group should be able to demonstrate two 
interoperable implementations of each feature.

For every part of the OWL DL specification we currently have multiple 
implementations that can implement it, and in fact for every subset 
of the language features that doesn't contain BOTH "inverseOF" and 
"OneOf" we have multiple working and interoperable implementations 
either complete at this time, or coming soon.
  So from a process point of view, we believe the current OWL DL 
implementation meets the requirements for advancement.

2 - Your second issue is a more important one, you suggest that for 
the current OWL DL, which includes  OneOf there is a lack of 
practical implementation experience -- this is not quite correct.  We 
have numerous implementations that support oneOf quite well.  What is 
correct is that for the whole of OWL DL, which includes BOTH 
owl:inverseOf and owl:oneOf there is a problem -- if both are used 
together, in the same ontology, there is a possibility that the 
solution to some queries will take a very long time (that is, that 
there is no effective algorithm).  This is true -- although the 
language containing both of those features is decidable (it is, and 
it's not just Ian's assertion, the proofs are quite well known and 
studied), the algorithmic complexity is potentially quite high. 
However, there are other solutions than removing owl:oneOf, for 
example removing owl:inverseOf would work as well.
  In fact, the problem is that there are two useful features in OWL DL 
(inverse properties and designators such as oneOf and hasValue) that 
are widely used in practice.  The WG believes that removing either 
one of these would cause there to be many valuable use cases that 
could not be represented in OWL DL.  However, there is no problem 
with documents that use either Inverses XOR designators, it is only 
when both are used together that the problem occurs.  A "common 
sense" analogy is found in medicine where there are many drugs that 
are each beneficial, but when used together they can cause bad side 
effects -- inverse and oneOf are analogous to two such drugs.

In your comment you write:


We find that the draft documents make it clear that OWL Full systems will
not have full reasoning support and that therefore users will not be too
surprised when there is a resulting migration cost from one OWL Full 
system to another.

so we believe that if the documents made clearer that using BOTH 
oneOf and inverseOf (and their various forms) could lead to an 
unexpected rise in complexity, we would set the expectation 
correctly. In that way the current OWL DL subset would be easier to 
understand, and the design rationale behind it better understood.

Thus, given these two below, we propose that the WebOntology working 
group will make the issue above clearer and will write text to appear 
in the Reference, S&AS and Test documents that explain the above. 
However we would also propose not to remove these very useful 
features from OWL DL or to have a long CR period mandated by this 
specific issue.

If this outline of a solution is acceptable to you, we will produce 
proposed text properly setting the expectations with respect to 
inverse and oneof, and request your approval before closing this 
comment.

We look forward to your response
   Jim Hendler
   WebOnt Co-Chair, on behalf of the Working Group







At 15:18 +0100 5/9/03, Merry, Martin wrote:
We wish to comment on the usefulness of OWL DL as a sensible subset of OWL
Full.

We're concerned that OWL users should have their expectations met when they
use OWL compliant systems.

We find that the draft documents make it clear that OWL Full systems will
not have full reasoning support and that therefore users will not be too
surprised when there is a resulting migration cost from one OWL Full system
to another.

We are concerned, however, that OWL DL is presented as a sensible stopping
point before OWL Full, where there are greater guarantees.

The theoretical results for the decidability of OWL DL are interesting but
not particularly helpful. OWL Lite is justified by practical results in DL
systems (primarily from Ian Horrocks). There is no such practical experience
for the OWL DL subset.  We would like to see such practical experience
before OWL exits candidate recommendation.

In particular, we would like to see adequate practical implementation
experience of the OWL DL constructs owl:oneOf and owl:hasValue.  We believe
that this should include the goal that OWL DL reasoners can make a
reasonable attempt at classic NP complete problems (such as the 3-SAT
problem and the subgraph isomorphism problem) which can be straightforwardly
encoded within OWL DL.  For example, any such problem that can be solved in
seconds by a specialised reasoner should be soluble by a general OWL DL
reasoner in minutes rather than years.

An alternative, would be to redefine OWL DL downwards, excluding owl:oneOf
and owl:hasValue, which would then be subject to the health warnings of OWL
Full - i.e. use of these constructs means that your ontology is likely to be
outside the limits of practical reasoning. Such a redefinition of OWL DL,
could sensibly accompany a redefinition of OWL Lite to exclude complete
class definitions.


Martin Merry
HP Semantic Web Programme Manager



Martin Merry
HP Semantic Web Programme Manager



-- 
Professor James Hendler				  hendler@cs.umd.edu
Director, Semantic Web and Agent Technologies	  301-405-2696
Maryland Information and Network Dynamics Lab.	  301-405-6707 (Fax)
Univ of Maryland, College Park, MD 20742	  240-731-3822 (Cell)
http://www.cs.umd.edu/users/hendler

Received on Friday, 9 May 2003 18:56:20 UTC