Message 12303 of the SUO list

Subject: Re: W3C approves RDF and OWL as recommendations
         (original subject header for the discussion but inappropriate here)
Date: Mon, 16 Feb 2004 19:10:24 +1000
From: Philippe Martin
In-reply-to: msg12300 by Danny Ayers (in answer to msg12299 by Asha)
Follow-up: msg12321 by Adam Pease (answered in msg12331 by Philippe Martin)

> > Asha:
> > 'Being a parent' is merely a relational property, while 'parentage'
> > (parenthood, but not the act or process of parenting) is a relation
> > of parent to child. Or, more generally, the relational property of
> > being a cause is just a monadic reduction of causality, the relation
> > of cause to effect. So, instead of classes and relational properties,
> > we have rather to speak of classes and their relationships as two
> > sorts of distinct entities.  
>
> Danny:
> Within OWL DL, yes. Within OWL Full (which encompasses 'generic' RDF)
> properties are classes too. ...


I do not think that Asha's main point (if I understood it) has been
answered. Given the additional precisions in http://www.eis.com.cy/
I think that Asha is referring to the fact that each "role-related
notions" such as "parent" (not the parenting process) may be represented
by a relation type (e.g. parent) and/or a concept type (e.g. Parent),
and that these two categories are exclusive (e.g. in the SUMO, parent
would be abstract while Parent would be physical), thus leading to
(i) an increase of the complexity of knowledge entering/management (since
the notions have to be represented twice and then connected by definitions
to permit inference engines to compare the representations using the 
relation types with the representations using the concept types), and/or
(ii) a decrease in knowledge sharing/retrieval possibilities (this is
actually the most likely outcome).

As a partial solution, in Ontoseek and WebKB-2, certain role-representing
concept types may be used in relation nodes, as in the following example
in CGLF: [Person:Tom]->(Parent)->[Person:Mary]. In WebKB-2, the constraints
are that Parent is a subtype of pm#thing_that_can_be_seen_as_a_relation
and that Mary has been declared as instance of Parent. In Ontoseek, if I
remember correctly, I think that Mary may simply be instance of Person
if Parent has been declared has a role type and is a direct subtype of
Person. Assuming that no relation type name parent exists in the ontology,
to export the above representation into KIF, such a relation type should
be generated and in its signature, the destination type should be Parent
or Person. Then, the KIF representation could be: (parent Tom Mary).

An extension of this approach that I will probably implement in the
medium-term is to permit a signature to be associated to a concept type,
e.g. specifying that if Parent is used as a relation, it must connect an
instance of Animal to another instance of Animal. E.g. in the FT notation:
  Parent (Animal,Animal) < Animal;
could be converted into the following in KIF+SUMO:
  (subclass Parent Animal)  (instance parent binaryRelation)
  (domain parent 1 Animal)  (domain parent 2 Animal)
  (=> (parent ?A1 ?A2) (Parent ?A2))
  (=> (Parent ?A2) (exists ((?A1 Animal)) (parent ?A1 ?A2)))

Such a shortcut can also be used in KIF, e.g. by defining the relation
roleTypeDomain in such a way that the following 3 formulas are equivalent
to the 6 previous ones:
  (subclass Parent Animal)
  (roleTypeDomain Parent 1 Animal)  (roleTypeDomain Parent 2 Animal)
Yes, that would involve string manipulation but KIF can do that.

A relation such as roleTypeDomain cannot be "defined" in RDF+OWL but it
could be added as a primitive to OWL. That would address one of the points
raised by Asha. However, in my opinion, that would not be sufficient for
RDF+OWL to be an adequate language for knowledge representation and 
sharing since 
(i) there are still many important constructs (related to contexts,
sets, and numerical quantifiers) that can only be represented in 
(very) ad-hoc ways or not at all (e.g. see the 2nd half of
http://www.webkb.org/doc/model/comparisons.html
i.e. starting from the section "contextualization"), and as John Sowa
pointed out, it is not the role of a "knowledge sharing" language to 
reduce the expressivity, it is up to each inference engine to select the
(parts of the) representations that it can use, and warn about the fact
that it ignored some information (or advertize its capabilities, as in
the OKBC protocol),
(ii) the model and the adopted conventions are low-level (i.e. they 
lead users to represent the same thing in many different ways, many of
which are uncomparable or very hard to compare automatically), prevent
very important checks (e.g. subsumption link cycles are allowed because
the used subtype links are not "proper subtype" links), and make the
conversion/presentation of the RDF representations into high-level 
languages and controlled Englishs quite awkward,
(iii) its XML syntax makes the detection of errors difficult (this
syntax is not supposed to be seen but in practice it is, if only by
researchers and those who develop RDF-related tools).

Philippe