Re: Changes to make S&AS consistent with RDF Semantics document

Thank you for your response to my question, Peter.
In this message I have some remarks and two other questions
for the S&AS editors (and the WG).

[...]

>> 
>> Both S&AS and RDF Semantics define datatype maps
>> to be partial maps from URI references to datatypes.
>> RDF Semantics assumes, in addition, that each datatype
>> map contains rdf:XMLLiteral.
>> S&AS assumes, in addition, that each datatype map
>> contains xsd:string and xsd:integer.
>> 
>> S&AS could be made consistent with RDF Semantics
>> by correcting the third definition in Section 3.1 
>> in for example the following way:
>> 

(*)
>> "As in RDF, a datatype map D is a partial mapping from
>> URI references to datatypes that maps rdf:XMLLiteral
>> to the built-in XML Literal datatype defined in the
>> RDF Concepts and Abstract Syntax document [RDF Concepts].
>> In addition, it is assumed that datatype maps
>> map xsd:string and xsd:integer to the appropriate 
>> XML Schema datatypes."

>
>I do not think that this change is appropriate, because rdf:XMLLiteral is
>not a required part of OWL.
>
>> A corresponding addition to the first definition in 
>> Section 5.2 would then be needed.
>> 
>> It should be noted that if XMLLiteral is not added
>> by default to each datatype map, as in the RDF Semantics,
>> document, then S&AS seems to be inconsistent.
>> To see this, consider the following example:
>> 
>> RDF graph G, just two triples
>>   v p l
>>   p rdfs:range rdfs:Literal
>> where l is an ill-typed XML literal.
>> Since G has no rdfs-interpretations, it has no D-interpretations
>> for any datatype map D, and also no OWL DL interpretations
>> for any datatype map D, so G is OWL DL inconsistent.
>> 
>> Abstract syntax ontology O, containing
>>   Individual(v value(p l))
>>   DataProperty(p range(rdfs:Literal))
>> If D is a datatype map that does not contain rdf:XMLLiteral,
>> then O is consistent.
>
>Agreed.
>
>> It is clear that the translation of O with the mapping T of 
>> S&AS contains the RDF graph G.
>> This contradicts the corollary to Theorem 1 in Section 5.4
>> in S&AS, for any datatype map D that does not contain
>> rdf:XMLLiteral.
>> (I am using here my first comment above, that Theorem 1 etc.
>> should be read as holding for a certain datatype map.)
>
>Agreed.
>
>This is an area that needs updating in OWL, now that RDF finally has
>workable datatypes.  The need for updating has been known for some time
>now, and wording was put in S&AS that S&AS depended on the details for 
RDF
>datatyping that had not yet been finalized.
>
>My suggestion would be to weaken the corollary, adding the condition that
>the corollary only holds when the datatype map includes the RDF mapping 
for
>rdf:XMLLiteral.  Other theorems would have to be similarly changed.

This change has been made, which solves my problem, in the sense that
the above error in the correspondence theorem no longer exists.
However, there still remain questions.
Whether you like it or not, rdf:XMLLiteral is part of OWL, in the sense
that each OWL DL consistency checker or OWL Full consistency checker that
is worthy of the name should recognize that the RDF graph G that I 
mentioned,
     v p l
     p rdfs:range rdfs:Literal
where l is an ill-typed XML literal, is inconsistent. 
Whether XMLLiteral is in or out of the datatype map D, OWL DL entailment 
as well as OWL Full entailment now always incorporates the meaning of 
rdf:XMLLiteral, inherited from the RDFS Semantics. 
 
In the light of this, what do you mean with your statement above:
>... rdf:XMLLiteral is not a required part of OWL.
?

By leaving out of Section 3 the assumption about XMLLiteral, in the 
current way, the abstract syntax-based, direct semantics is not 
just another side of the same coin as the OWL DL semantics.
It looks like the current editor's version of S&AS has three inequivalent
versions of OWL semantics: OWL Full, OWL DL, and a third version that 
might
be described as 'abstract syntax OWL without XMLLiteral'.
Each of these versions is labeled normative in S&AS.

Doesn't this contradict the WG decision about semantic layering, to have
two semantics versions, now called OWL Full and OWL DL?
One way to solve this would be to restrict the definition now in 
Section 5.4 of OWL DL entailment to datatype maps with XMLLiteral,
and to state another definition of OWL DL entailment for datatype maps 
without XMLLiteral, in terms of the direct semantics and the mapping T,
as this is a known case of conflict between the direct semantics
and OWL DL semantics.
This would realize the desire that XMLLiteral is not required in
OWL DL.

Another, simpler, way to rule out the third version of the semantics 
would be to adopt the suggestion (*) that I give above, in Section 3 
of S&AS.
This is not to say that the changes should then be undone to the 
formulations in Section 5 of the entailment definitions and the 
correspondence theorem etc., I like them to be (and actually asked 
to make them) completely explicit, in the current way.

===


>> 
>> (Section 3.1: second bulleted condition:)
>> >-It is now assumed that LV contains each Unicode string
>> >and each pair of two Unicode strings.
>
>This should actually be weakened to pairs of Unicode strings and language
>tags, or whatever the RDF model theory says.  Again this is something 
that
>has undergone recent change in the RDF model theory.
>
>> >For the correspondence with Section 5, it would be
>> >sufficient to assume only that plain literals in
>> >V (and L) are contained in LV.
>> On further reflection, it seems that not only the assumption
>> about plain literals but also the assumption about
>> typed literals could be weakened.
>> The condition could be rephrased, for example, as follows:
>> "LV, the literal values of I, is a subset of R that
>> contains the values of plain literals in V, and,
>> for each datatype d in D and well-typed literal
>> "v"^^d in V, the value L2V(d)(v)."
>
>Part of this would work, except that it would have to refer to the RDF
>model theory to pick up the ``value'' for literals. 

This was indeed a slip.  Ian and I corrected it to become 
essentially the following condition:
(**) "LV, the literal values of I, is a subset of R that
contains the values of plain literals in V, and, for each
datatype d in D and v in  L(d) intersect V,  the value
L2V(d)(v)."

After discussion with me, Ian adopted this definition in the 
"PR version of S&AS" of Tuesday (with a corresponding phrase 
in the beginning of Section 5).
Yesterday, Peter sent a message to webont stating that this
would change the OWL Full semantics and without further 
discussion, within one hour, Peter, from outside the WG, 
put the earlier condition back in, which was subsequently 
confirmed by Ian.
That this would change the semantics, is indeed an objective
reason for undoing the change.
I can live with the current version of this definition. 
It does not seem to be inconsistent with the RDF Semantics 
document.
(However, nobody has done a complete review of the proof
appendix of S&AS, which might uncover such an inconsistency.)

The reason for proposing (**) was that a similar condition
in the RDFS semantics led to a "great simplification" [1]
of the proof theory, although this occurred in a situation
without datatypes, apart from XMLLiteral.

[...]

Herman ter Horst

[1] http://lists.w3.org/Archives/Public/www-rdf-comments/2003OctDec/0161.html

Received on Thursday, 4 December 2003 11:43:03 UTC