Re: Returning to OWL and application profiles

On Tue, Oct 12, 2010 at 4:08 PM, Mark van Assem <mark@cs.vu.nl> wrote:
> Hi Mikael,
>
> I just defended my PhD thesis [1] last week and it contains a section 7.5
> devoted to APs specified in OWL (also referring to your AP constraint
> language). I suggest you have a read :-)
>
> What I propose is to create subclasses of a particular class such as
> lib:Book, e.g. my:Book and constrain property values on my:Book so that no
> inconsistency arises with his:Book (which may have entirely different
> constraints).

This approach seems very good from the modelling point of view, but
I'd like to ask wether it is realistic in a pragmatic Linked Data
world.

If I want to add specific constraints on, for instance, dct:creator,
and I create my:creator, I will reach interoperability with others
using dct:creator only through inferencing. This doesn't seem very
straightforward to me. We are lacking actual use of Linked Data today,
and I feel that adding more complexity in the data model is likely to
create more barriers.

As librarians sitting on a big mass of data, our immediate need is to
be make our data understandable in the global Linked Data world, and
I'm not sure that encouraging us to systematically create our own
classes and properties rather that reusing existing ones, for the sake
of expressing patterns, is the right way to go. I see a risk to
encourage the creation of a lot of redundant, slightly different but
almost similar vocabularies (we already have 3 flavours of FRBR out
there...)

What I like in the application profile approach is the idea that I can
reuse *existing* classes & properties, and at the same time, express
the pattern that my community should reuse for describing similar
resources.
It's actually 2 separate needs :
- a need for visibility and interoperability, that can be fulfilled by
existing vocabularies with their OWL semantics that are consistent
globally
- a need for describing domain-specific patterns.


We call this "collection-specific value ranges" (with
> collections referring to cultural heritage collections), but cardinality
> constraints is basically the same story.
>
> The only thing that then has to be added is a way to classify every book as
> either my:Book or something else (the "something elses" representing wrongly
> specified books, i.e. violations of the AP constraints). This is probably
> abusing OWL a bit, but it can probably be done.
>
> An alternative is to just accept OWL as a "syntax" for the AP constraints,
> and implement your own checker on top of that. This removes the need to
> develop your own language (and parser) which will contain almost the same
> syntactical elements anyway.

I like the idea that the AP should be something that could be
implemented following different syntaxes, maybe including OWL, but not
excluding other approaches that wouldn't make it mandatory to declare
local properties and classes systematically when additional semantics
or constraints are needed.

Emmanuelle
>
> (As far as I can tell this is also what DC-APs are intended to do. I had
> discussions with Tom Baker about this, and he was part of the committee that
> accepted my thesis last week so apparently I didn't write rubbish :-)
>
> Best,
> Mark
> [1]http://www.cs.vu.nl/~mark/papers/thesis-mfjvanassem.pdf
>
> On 12/10/2010 13:50, Mikael Nilsson wrote:
>>
>> Hi Antoine,
>>
>> tis 2010-10-12 klockan 13:11 +0200 skrev Antoine Isaac:
>>>
>>> Hi Mikael,
>>>
>>> Thanks for starting this interesting thread :-)
>>
>> Thanks for an interesting reply :)
>>
>>>
>>> Well, I have what I think is a quite traditional SW background, and I'd
>>> be tempted to turn the argument the other way round ;-)
>>> If the instances of one class in an AP have one title and the instances
>>> of that class in another AP have two titles, then it is perhaps that these
>>> two APs are thinking of two different classes, really. Possibly they could
>>> be two subclasses of a common superclass, but two different classes, still.
>>> I think this is quite in line with Jeff's message suggesting you could do
>>> things like this
>>> baz:Widget rdfs:subClassOf foo:Widget.
>>> to make the commitment of your various APs a bit clearer.
>>
>> Yes, I do believe that conceptually, most APs can be described using the
>> trick of the creation of subclasses, even though the subclass may well
>> consist of the same individuals (!)
>>
>>>
>>> I'm also a bit puzzled by your "APs define domain-specific structural
>>> constraints while OWL adds semantics to existing classes." Again, I am
>>> pretty new to the APs as practiced in the DC realm, but why wouldn't you
>>> consider classes (and APs built on top of them) from a (SW) semantic
>>> perspective?
>>> I understand that OWL and RDF will fail for arrangement/presentation of
>>> data (order of XML elements, e.g.), which is not really about semantics. But
>>> to me--and to many in the SW community--cardinality belongs to semantics of
>>> classes and properties.
>>
>> Let me give a simple example. Let's assume I am designing a simple REST
>> service for retrieving metadata records from a library repository (never
>> mind the 303s etc).
>>
>> I decide to return Turtle representations of the form
>>
>> myrepo:book123 rdf:type lib:Book,
>>                dct:title "Moby Dick",
>>                dct:creator myrepo:author345 .
>>
>>
>> I can define this pattern using OWL restrictions on the lib:Book class.
>>
>> Later on, I want to define an extended API for use by library partners.
>> This API returns records of the form
>>
>> myrepo:book123 rdf:type lib:Book,
>>                dct:title "Moby Dick",
>>                dct:creator myrepo:author345,
>>                lib:numCopies "5".
>>
>>
>> This record describes the exact same individual but using a different
>> application profile.
>>
>> To solve this with the subclass method I would need to define a subclass
>> lib:ExtendedBook that captures the additional property, but has the same
>> extension.
>>
>> My point is that application profiles describe metadata *patterns*,
>> independently of the semantics. None of the above properties are likely
>> required (cardinality>  0) for the Book class in any case. And there may
>> well be multiple such patterns for the same class or set of things.
>>
>> In the DC world, the use of application profiles is often framed in this
>> way - an application profile does not describe a class, it describes the
>> metadata records emitted or accepted by a system (hence the term
>> "application").
>>
>>>> But that's exactly it - the semantics is "alternative" and on its face,
>>>> the semantics of the published OWL file is something else entirely.
>>>>
>>>> As a concrete example, if I serve an OWL file from my web server, using
>>>> application/rdf+xml as suggested by the OWL specs [1], the
>>>> interpretation will be as RDF triples using the RDF and OWL built-in
>>>> semantics, thus resulting in the generation of new triples, potentially
>>>> contradiction with other ontologies, and not in validation as expected.
>>>
>>>
>>> Well, contradiction (aka inconsistency) is what is used for data
>>> validation in the RDF/OWL world. So if we can detect contraditions I
>>> wouldn't be unhappy, at least from the perspective of us able to
>>> validate some data :-)
>>
>> My point is that the two ontologies needed to describe the two uses of
>> lib:Book above would be contradictory. One would the be left wondering,
>> "what is the semantics of the lib:Book class, is it as defined by
>> ontology A or B?" and that's not good. The truth is that the semantics
>> of lib:Book is less constrained (all the above properties optional), and
>> the constraints appear only to describe certain *records*, not the
>> lib:Book class itself.
>>
>>>
>>> It's really a situation where the cons for an OWL approach (and being
>>> awkward to manipulate is not the least default OWL has, sure) could be
>>> balanced by some strong pros, and we should not discard them too
>>> quickly!
>>
>>
>> I agree that OWL has advantages, but we need to develop an understanding
>> of what kinds of validation we can achieve.
>>
>> Maybe we arrive at the conclusion that we need to accept a different set
>> of use cases for OWL-based ontologies. I'm not very happy with that
>> thought, however.
>>
>> I therefore lean towards a solution based on true syntactical constraint
>> language on RDF graphs. Alistair Miles did some experiments in that
>> direction, but they seem to have disappeared from the net.
>>
>> It would be extremely interesting to try to figure out whether there are
>> significant use cases for purely syntactic constraints in the LD domain.
>>
>> /Mikael
>>
>>
>>
>
>

Received on Tuesday, 12 October 2010 15:09:46 UTC