Re: OWL-S Reasoning -- Instances vs. Classes

On Feb 13, 2004, at 3:55 AM, Michael Lutz wrote:

> as we did not get any responses to our first (lengthy) e-mail, here's 
> another (much shorter) try.

Sorry, lots going on.

> Reasoners like RACER are able to compute taxonomies between (OWL) 
> concepts (ontology classes). As OWL-S itself is an ontology of 
> services

It's an ontology about services. Maybe of services if we consider the 
tripartite division of Atomic, Composite, and Simple, I suppose. What I 
meant is that we have no specific service types defined.

>  - how is it possible to reason about OWL-S service descriptions 
> (which are instances in that ontology) in order to derive taxonomies 
> of services?

In Processes as Classes (PAC), there was the thought that of any two 
services, you could derive whether one was more general than the other 
by computing their subsumption relation. However, it's not clear that 
that's useful or that it's even common that services defined as 
concepts in this matter *would* tend to have interesting subsumption 
relations. This suggests that there was no loss, at least ;)

In Processes as Instances (PAI), there is, I think, the intuition that 
"Service type" and "Service definition" are somewhat separate. If you 
want a hierarchy of service types based on the definitons, you need to 
define some classes which have as members those definitions. We have to 
define some classes just to get some of expressivity we want (e.g., 
that a class has exactly 3 inputs), so there should be emerging some 
(uninteresting?) class hierarchies). You might be able to generate 
classes by looking at these definitions and thus recapture some of the 
PAC thought, but I'm not sure this is beneficial.

Reasoners with ABox capability (like Racer and Pellet) offer a number 
of services for dealing with indiviudals (Pellet, for example, offers 
sound and complete (if slow and a bit limited) conjunctive Abox query). 
I think there is the thought (:)) with PAI that this sort of thing is 
more useful in the arbitrary case. For any particular purpose, you 
should define your own classes of processes and do instance checking 
(or retrieval, or realization) against the fully classified class tree.

I don't know if this answers your question, but even if not, I hope it 
helps :)

What kinds of problems do you want to solve with such taxonomies?

Cheers,
Bijan Parsia.

Received on Friday, 13 February 2004 07:25:07 UTC