Re: ISSUE 131 (OWL R Unification): Different semantics on syntactic fragment

On 14 Aug 2008, at 23:18, Jim Hendler wrote:

>
> On Aug 14, 2008, at 5:46 PM, Ian Horrocks wrote:
>>
> [snip]
>
>>
>> Actually, I much prefer your idea of defining conformance w.r.t.  
>> the syntactic fragment, i.e., conformant OWL RL reasoners must be  
>> complete for query answering as defined in the current section  
>> 4.4. This would be a slightly less strict condition, but it has  
>> the advantage of being non-procedural and of coinciding with the  
>> syntactically defined class of ontologies for which rule-based  
>> implementations are complete.
>
> I was with you (and Michael) up to here - but now I get worried.   
> The real problem is that while this is easy to say, most  
> implementors don't have PhDs in AI or strong enough logical  
> backgrounds to prove that their implementations are "complete for  
> query answering as defined in section 4.4" --  certainly a lot of  
> people may well write programs in either procedural languages or  
> using logic programming which would likely be sound and may well  
> cover the cases - but they'd have no way to prove it (and as the  
> code may be proprietary or part of a much larger system, there'd be  
> no easy way to "outsource" the proof).    Seems to me this could  
> have one of two bad results
>  1 - people will simply ignore the definition and claim  
> conformance, which is bad, or
>  2 - people will take it seriously, which will keep them from  
> building the easy rule-based implementations they could have by  
> simply using their favorite rule systems (and integrating that into  
> whatever application they are using), thus hindering the growth of   
> the market for OWL ontologies)

What I should perhaps have mentioned is that the section 4.4 stuff is  
actually part of an explicit statement that implementations of the  
rule set *are* (sound and) complete for the defined query answering  
entailments, so they are conformant by definition. So no PhD or logic  
background required, although we should, or course, make sure that  
people with logic backgrounds have done the relevant proving for us :-).

>
> For OWL 1.0, we realized that there was no easy way to deal with  
> this, and we developed the idea of having soundness as a stated  
> goal and a test suite -- the more tests you passed, the more likely  
> you were to be correct w/respect to the semantics - not a great  
> solution, but it worked pretty well in its day.   Problem is this  
> WG isn't doing a test suite as far as I can tell, so we cannot use  
> this out.

In Test Cases [1] we defined an "OWL consitency checker", and also  
defined what it means for such a consistency checker to be complete  
and terminating. Our definition was actually quite strict:  
consistency checkers can answer Consistent, Inconsistent or Unknown,  
and they "SHOULD NOT return Unknown". A consistency checker must be  
sound, i.e., "MUST return Consistent only when the input document is  
consistent and Inconsistent only when the input document is not  
consistent". A consistency checker is complete and terminating if  
"given sufficient (but finite) resources (CPU cycles and memory) and  
the absence of network errors, it will always return either  
Consistent or Inconsistent".

We then used these basic definitions, along with the relevant syntax/ 
semantics specs, to define consistency checkers for each of OWL Lite,  
DL and Full. We also explicitly defined a complete OWL Lite  
consistency checker, being one that is complete and terminating as  
above. We didn't explicitly define a complete OWL DL consistency  
checker because, at that time, we didn't have a practical decision  
procedure or any sound and complete reasoners for OWL DL. We didn't  
explicitly define an OWL Full consistency checker because the  
language is known to be undecidable.

What has been proposed, e.g., at the recent F2F, was to have  
something similar in the OWL 2 spec. For OWL 2 RL, however, we need  
something a bit different along the lines we have been discussing here.

Ian

[1] http://www.w3.org/TR/owl-test/


>
> So it seems to me the WG could take a hard line on what conformance  
> is, at the risk of it being mainly ignored (and I can tell you from  
> personal experience that going out into the blogosphere and saying  
> that some company pushing a product isn't conformant is not a smart  
> thing to do if you value your sanity) or we can take an easier line  
> and at least get people to think about the issue of soundness,  
> which is easier to assert (although not necessarily to prove)
>
> My personal opinion is to take the easier line, because OWL 1.0 has  
> shown us that the market is actually pretty good at working these  
> things out -- a lot of OWL DL reasoners started to emerge, but  
> some, like Pellet, Fact and Racer, by dint of being well grounded,  
> became widely used, while a lot of "partial implementations" (many  
> of which were actually rule-based) didn't make it.
>
> Heck, I have a PhD in AI, but it's a lot of years out of date with  
> respect to the knowledge I'd need to prove conformance as defined  
> in the quote above -- so perhaps it would be better to stick to  
> things that us "simple folk" out here can understand
>   -JH
> p.s. an alternative is to bring back the test suite, but that is a  
> lot of work as the OWL 1.0 folks can attest to!
>
>

Received on Friday, 15 August 2008 17:19:41 UTC