Re: Performance issues with OWL Reasoners => subclass vs instance-of

On Fri, 15 Sep 2006, Phillip Lord wrote:

>
>
>
>
>>>>>> "WB" == William Bug <William.Bug@DrexelMed.edu> writes:
>
>  WB> CLASSes represent UNIVERSALs or TYPEs.  The TBox is the set of
>  WB> CLASSes and the ASSERTIONs associated with CLASSes.
>
>  WB> INSTANCEs represent EXISTENTIALs or INDIVIDUALs instantiating a
>  WB> CLASS in the real world.  The ABox is the set of INSTANCEs and
>  WB> the ASSERTIONs associated with those INSTANCEs.
>
>
>
> I'd take a slight step back from this. You can think of classes and
> instances in this way. But in the OWL sense, a class is a logical
> construct with a set of computational properties. "Instances" is a
> more difficult term. OWL actually has individuals. The instance store
> uses "instances" because they are not really OWL individuals.
> There is also a philosophical concept of what a class is, what a
> universal is an so on, which may be somewhat different, and is also
> open to debate.
>
>  WB> Properly specified CLASSes are defined in the context of the
>  WB> INSTANCEs whose PROPERTIES and RELATIONs they formally
>  WB> represent.
>
>  WB> Properly specified INSTANCEs are defined via their reference to
>  WB> an appropriate set of CLASSes.
>
> Think this would be circular. An OWL class is defined by the
> individuals that it might have in any model which fits the
> ontology. Not just the individuals it has an a specific model.
>
>
>  WB> Reasoners (RacerPro, Pellet, FACT++) generally have
>  WB> optimizations specific to either reasoning on the TBox or
>  WB> reasoning on the ABox, but it's difficult (i.e., no existing
>  WB> examples experts such as Phil and others can cite) to optimize
>  WB> both for reasoning on the TBox, the ABox AND - most importantly
>  WB> - TBox + ABox (across these sets).
>
> ABox is more complex than TBox, although I believe the difference is
> not that profound (ie they are both really complex). For a DL as
> expressive as that which OWL is based on, the complexities are always
> really bad. In other words, no reasoner can ever guarantee to scale
> well in all circumstances.

Once again: pure production/rule-oriented systems *are* built to 
scale well in *all* circumstances (this is the primary advantage they 
have over DL reasoners - i.e., reasoners tuned specifically to DL 
semantics).  This distinction is critical: not every reasoner is the same 
and this is the reason why there is interest in considerations of using 
translations to datalog and other logic programming systems (per Ian Horrocks suggestion below):

>Another interesting approach that has only recently been presented by
>Motik et al is to translate a DL terminology into a set of disjunctive
>datalog rules, and to use an efficient datalog engine to deal with
>large numbers of ground facts. This idea has been implemented in the
>Kaon2 system, early results with which have been quite encouraging (see
>http://kaon2.semanticweb.org/). It can deal with expressive languages
>(such as OWL), but it seems to work best in data-centric applications,
>i.e., where the terminology is not too large and complex.

I'd go a step further and suggest that even large terminologies aren't a 
problem for such systems as their primary bottleneck is memory (very 
cheap) and the complexity of the rule set. The set of horn-like rules 
that express DL semantics are *very* small.

Chimezie Ogbuji
Lead Systems Analyst
Thoracic and Cardiovascular Surgery
Cleveland Clinic Foundation
9500 Euclid Avenue/ W26
Cleveland, Ohio 44195
Office: (216)444-8593
ogbujic@ccf.org

Received on Friday, 15 September 2006 15:42:32 UTC