Re: A semantic layering proposal

Obviously, I don't like this idea, but I can see that it has some 
merit as being at least a coherent point of view. I have one overall 
comment and a strong technical objection to one of the sub-proposals.

The overall comment is that if this proposal is adopted, then OWL 
will be an excellent notation for using XML to send data to a DL 
reasoner, which will be a useful, if modest, contribution to the 
growing battery of software tools available. However, that is all it 
will be. It will not, in particular, be a useful semantic web 
ontology language. I think that to adopt this kind of proposal, in 
which the operational aspects of the language are completely 
determined by the need to preserve the integrity of the abstract 
syntax, ignoring or overriding all issues of interoperability with 
other notations and sources of information, would be a failure of our 
collective imagination and would not significantly contribute to the 
creation of the SW. This is not a total disaster, of course, since it 
simply means that others, with more imagination and daring, will make 
things happen. But it will be a missed opportunity. One thing that 
will certainly happen is that people will use the OWL vocabulary 
inside RDFS without conforming to the abstract syntax: it is just too 
useful to ignore, and the many restrictions of the abstract syntax 
are too tiresome and burdensome to handle. In fact, many of the 
people on the WG will do this. Some of them already are doing it, and 
we cannot prevent it. (What are we going to do, shoot them?) So we 
have two options, seems to me: tell them they are doing something 
illegal, or offer them some guidance. I would rather be a missionary 
than a policeman.

>The WG has for some time been wrestling with difficult problems
>surrounding the issue of semantic layering. The basic problem arises
>from the observation that layering OWL on top of RDF in the DAML+OIL
>style leads to an incorrect set of entailments.  Proposed solutions
>have included asking RDF core to provide a mechanism for switching off
>the RDF meaning of some triples (dark triples) and weakening the
>semantics of OWL such that many intuitive entailments are lost (e.g.,
>(A or B) not entailing (B or A)).
>
>More recently solutions have involved restricting the set of RDF
>graphs that are considered to be valid OWL, and only providing an OWL
>semantics for these graphs. This proposal follows the same line,
>restricting the set of OWL/RDF graphs to those that can be generated
>from the OWL abstract syntax.
>
>The proposal is as follows:
>
>* The abstract syntax document provides the basic definition of OWL
>   and specifies the subset of possible RDF graphs that are valid OWL
>   graphs (i.e., those that can be derived via a mapping from this
>   syntax) [1].
>
>* An XML presentation syntax can be developed from the RDF triple
>   mapping.
>
>* RDF triples conforming to mappings from the abstract syntax are the
>   normative exchange format for OWL ontologies [2].
>
>* The normative semantics for OWL is given by [3], which defines the
>   semantics in terms of the abstract semantics. A definition of the
>   semantics w.r.t. the resulting RDF triples is also available [4],
>   but should be considered non-normative.
>
>* The link between OWL and RDF semantics is defined by [4]. What this
>   basically says is that *IF* RDF graphs are constrained to those
>   generated by the abstract syntax and *IF* extra semantic conditions
>   are added to the RDF MT, *THEN* entailment in this extended RDF is
>   the same as OWL entailment. (This is roughly what Pat has been
>   calling fast OWL.) If the extra semantic conditions are not
>   considered, then RDF entailment is still consistent with OWL
>   entailment, but it becomes a subset of the entailments justified by
>   the OWL semantics.
>
>* A form of "classes as instances" can be achieved by a naming
>   convention or external mapping that identifies pairs of classes and
>   instances. This is similar to techniques already being used in
>   implemented RDF engines. It does not support all the entailments
>   that would be possible by eliminating the distinction completely,
>   but it does allow for a layered architecture with complete OWL
>   entailment within each layer.

I strongly object to this. (Incidentally, I can see no reason why an 
implemented RDF engine would need to use punning, unless it were 
restricted in some way by a non-RDF class convention in some 
underlying VM.) Allowing 'punning' is a fundamental logical error 
which has many dangerous semantic repercussions, since names no 
longer have a well-defined denotation. In some ways, in fact, one 
could think of the use of the IEXT mapping in the RDF MT as a way of 
incorporating the utility of this device within a coherent model 
theory; but in that light, part of the basic utility of having 
circularities in class membership is that it does NOT require the 
user to resort to such crude, confusing and dangerous tricks in order 
to state simple conditions such as [rdfs:Class rdf:type rdfs:Class.] 
The problem here is not in either the 'naive' intuitions that such 
assertions make sense (they do) nor in the use of intensional classes 
in a semantic theory (which has a long and eminently respectable 
intellectual history behind it) nor even in the requirements of an 
efficient implementation, but rather in a mistaken but pervasive 
feeling that anything that steps outside a restricted layered 
universe is unfamiliar and dangerous. For inheritance class 
heirarchies in programming languages it may indeed be a non-no, but 
for semantics is is harmless and useful; but allowing punning IS 
dangerous, and would effectively make it impossible to use OWL safely 
with any other language. So this trades a very small advantage for a 
very big disadvantage, and as far as I can see to no real purpose. It 
is trivial to define a version of fast OWL which could allow classes 
as instances (relax the condition that IOC and IOT have to be 
disjoint) but otherwise conforms to the restrictions required for DL 
reasoners, and I think that there would be very little work required 
to map it into a suitable form. (One way to view it would be that the 
mapping would do this punning automatically, but invisibly to the 
user, rather than forcing people to 'layer' their thinking in order 
to express intuitive content. More ambitiously, it may be that DL 
style reasoning can be extended to the CL notation for FOL without 
too much trouble. But in any case, we should not impose semantically 
dangerous and intuitively confusing mechanisms on users simply to 
keep the language within some deductive complexity class.

>
>* A subset of OWL (OWL Lite) will also be defined [1,5].
>
>
>The proposal provides for a complete language specification, almost
>all of the elements of which are already available. It addresses all
>of the following issues:
>
>5.2 Language Compliance Levels
>5.3 Semantic Layering
>5.10 DAML+OIL semantics is too weak
>5.17 XML presentation syntax
>5.19-Classes-as-instances
>
>and proposes solutions that, it is hoped, will be acceptable to
>most/all of the WG.
>
>Regards,
>
>Peter Patel-Schneider
>Ian Horrocks

Finally, this proposal does have one overall great benefit to me, 
which is that since it consists entirely of documents written by 
other people, I needn't do any more of this tiresome and unrewarding 
work and can retire to my other pursuits.

Pat


>
>[1] http://www.w3.org/TR/2002/WD-owl-absyn-20020729/
>[2] http://www.w3.org/TR/2002/WD-owl-ref-20020729/
>[3] http://www-db.research.bell-labs.com/user/pfps/owl/semantics.html
>[4] http://www-db.research.bell-labs.com/user/pfps/owl/embed.html
>[5] http://www.w3.org/TR/2002/WD-owl-features-20020729/


-- 
---------------------------------------------------------------------
IHMC					(850)434 8903   home
40 South Alcaniz St.			(850)202 4416   office
Pensacola              			(850)202 4440   fax
FL 32501           				(850)291 0667    cell
phayes@ai.uwf.edu	          http://www.coginst.uwf.edu/~phayes

Received on Friday, 27 September 2002 22:56:40 UTC