layering (5.3, 5.10): Sardinia compromise?

There have been numerous conversations between WG members in Sardiia this
week.
Some have been about layering.
My take is that we have been arriving at a potential compromise.

The goal of the compromise is to postpone the layering issues rather than to
solve them.
i.e. we find a solution that is good enough for OWL 1.0; which satisfies
nobody, but doesn't go over anyone's threshold.

In essence the compromise would be to make OWL a weak semantic extension of
a subset of RDF.

I will unpack that a little and then give an example:

RDF defines a large set of RDF graphs; for OWL we syntactically define a
subset W (or alternatively its complement). OWL then only applies to graphs
in W. We choose fairly arbitrarily what goes in W and what doesn't. The main
idea is the things that we can agree on go in W, and those things we
disagree about don't.

A possible characterisation of  W is that a graph is in W if there is an
unproblematic mapping of it into some DL syntax. A possible characterisation
of the complement of W is that a graph is in the complement of  W if it
contains any bad triples, where a bad triple is one that doesn't really fit
into a standard description logic framework.

The "weak semantic extension" is intended to mean, that as long as we
restrict ourselves to W then whenever g rdfs-entails h, then g owl-entails
h. However we are not requirng any other relationship between the semantics.
In particular if properties in the domain of discourse or type as a property
cause problems for OWL semantics then we do not require OWL semantics to do
that just because the RDF model theory is constructed like that. We only
require the externally visible behaviour (i.e. entailment) to conform with
RDF.

Finally an example:

consider the one triple graph:

g0:
eg:a eg:p eg:b .

This is entirely unproblematic, and there is a wide consensus in the group
as to its meaning. Thus it is in W.
(A syntactic characterisation of those conditions is needed).

Under RDF there are a number of other one triple graphs that are entailed

g1:
_:x eg:p eg:b .

g2:
_:x eg:p _:y .

g3:
eg:a eg:p _:x .


g4:
eg:p rdf:type rdf:Property .

g5:
rdf:type rdf:type rdf:Property .

Now, g1, g2 and g3 are also unproblematic and so in W, hence g0 is required
to owl-entail g1, g2 and g3.
However, g4 and g5 may be problematic, since the first assumes that
properties are in the domain of discourse and the second assumes that
class-membership is in the domain of discourse.
If we choose not to have these things then we ensure by the definition of W
that g4 and g5 are not in W.
Then the question does g0 owl-entail g4 becomes syntactically ill-formed,
and we don't have to answer it. This is a good thing because we don't agree
on the answer.

This compromise would involve indicating to the coordination group that we
had postponed the issue, and they would need to ensure the relevant people
do continue to move the layering problem forward; but after OWL 1.0 is
cooked.

Jeremy

Received on Friday, 14 June 2002 09:22:12 UTC