Re: Web Rule Language - WRL vs SWRL

On 28 Jun 2005, at 20:09, Michael Kifer wrote:

>
>>
>> Sorry to be late joining the party (I take a vacation and world war
>> three breaks out!), but I feel compelled to add my "twopennyworth" (as
>> we say). Much of this has been said before (by Jim, Bijan and Holger),
>> but I will try to summarise...
>
> Hi Ian,
> welcome to the party :-)
>
>
>> Regarding DLP under Horn (equivalently DL/FOL) and LP semantics, it is
>> true that *under certain assumptions* the two semantics are
>> indistinguishable, but this does *not* mean that they are equivalent.
>> The assumption being made in this case (as Bijan has pointed out) is
>> that only ground entailment will be considered. This is a rather  
>> strong
>> assumption: it is common to compensate for weak representation
>> languages (such as DLP) by providing much more expressive query
>> languages (e.g., the case of SQL), and with more expressive languages
>> we can easily distinguish two different semantics (as illustrated in
>> [1]).
>>
>> This is particularly important in a layered architecture, where the
>> increased expressive power of higher layers will naturally allow the
>> two different semantics to be distinguished. This was the main point  
>> of
>> [1], i.e., that we should not fool ourselves about the degree of
>> interoperability that would be provided by the architecture proposed  
>> in
>> the so-called "updated layer cake". Such interoperability only exists
>> so long as neither tower extends above the DLP level *and* so long as
>> we only allow for very weak query languages; DLP does not, therefore,
>> provide a suitable foundation for a layered architecture, as higher
>> layers (or more expressive query languages) might make false
>> assumptions about the intended semantics of DLP ontologies.
>
> This wasn't clear from the paper.
> In any case, the claimed interoperability doesn't extend to the more
> powerful languages.

Exactly. So it is inappropriate to show LP and FO language towers  
interoperating via a common DLP base - they would sink in and topple  
over.

>
> Your argument is well-taken. However, if you are querying an ontology  
> with
> an LP language then you expect that CWA will be applied in the current
> state.
> That is, if your DLP ontology says that John has one child, Bill, and
> nothing else, then your query "get all people who have exactly one  
> child"
> (a query like the one in your paper) is expected to return John, since  
> in
> the current state he is not known to have more children.

This is the exact point - LP semantics would lead to an entailment that  
is *not* supported by DLP (which, as you will see from the paper, has  
FO semantics). Moreover, the ontology in question is actually an RDF  
ontology (as well as being a DLP one), so this is an example of LP  
semantics leading to an entailment that is *not* supported by RDF  
semantics (because RDF semantics would allow for other models in which  
John has additional children). Thus we can see that LP *cannot* be  
layered on top of RDF.


> There will be certain amount of semantic mismatch, but this most likely
> won't matter. From talking to people I get the impression that
> they build very simple ontologies using OWL. Most of them are in DLP  
> with
> some use of cardinality restrictions. However, when I press, it turns  
> out that
> people don't understand what cardinality restrictions mean in OWL. They
> think they are like cardinality constraints in databases.
>
> People think databases but use OWL.
> Now, this is a real semantic mismatch: people mean (and want!) one  
> thing,
> but get a completely different thing.

You need to get out more. I meet and interact with many users who are  
building large and complex ontologies, and who have a very good grasp  
of the semantics of OWL. Hopefully you caught the presentation from  
Christine Golbreich at the rules workshop and have been keeping up with  
the work being done in the SWBP working group. You should also look at  
what is being done by NCI, SNOMED and the Gene Ontology Consortium  
amongst many others.


>
>
>> Regarding the existing layered architecture of OWL, and the proposed
>> extension to SWRL, nobody (well not me anyway) claims that it is
>> perfect, and there will (as Holger has pointed out) no doubt be some
>> compatibility issues between different tools, but the situation is
>> hardly comparable to the one that is being proposed in the "updated
>> layer cake": as we have seen, even the very limited degree of
>> interoperability suggested by this diagram turns out to be a chimera.
>> In contrast, RDF, OWL and SWRL share a common semantic framework, and
>> allow for a relatively clean layering: OWL-Lite, OWL-DL and SWRL are
>> layered on a subset of RDF, but share the same semantics (and it can  
>> be
>> syntactically determined when RDF ontologies are within this subset).
>
> First of all, I disagree that RDF has the same semantics. It is just a
> language of facts. It all depends on the query language that you use  
> on top of
> RDF.  (E.g., the relational model can be queried with SQL or  
> first-order
> logic---two languages with very different semantics.)

I said that they share a common semantic framework that allows for a  
relatively clear layering. As we have seen, this is not the case for  
RDF/DLP/LP.

>
> Second, DLP is not the only interoperability option. A more general  
> one (and
> the one that I happen to like) treats the different languages as black
> boxes, which can be queries according to the native semantics of those  
> boxes.
> I was actually surprised that you referenced Eiter et al. paper,  
> because in
> my view they argue in favor of the 2-stack architecture with a defined
> interoperability layer based on the black box idea.

Now we are getting somewhere! Developing a framework that allows for  
multiple language types to co-exist in the semantic web while  
maximising interoperability is exactly what myself (and others) have  
been arguing for! I believe that the Eiter work, as well as other work  
by Motik, Rosati, Franconi, and others, provides a promising basis for  
such a framework, but that some of the details still need to be worked  
out.


>
>
>> Regarding the complexity of SWRL, yes it is a complex language, but as
>> seen from presentations at the rules workshop (and as pointed out by
>> Holger), some users *need* this level of complexity. For them, it is
>> essential to have a more expressive language that layers on top of  
>> OWL.
>> Should we thus forge ahead with the development of *two* fundamentally
>> incompatible rules languages? Surely (as Jim has pointed out) we can  
>> do
>> better than that, even if it requires a little more research effort!
>
> SWRL is complex in a wrong way. How many people use it in a way
> where programs come out substantially non-Datalog? (I.e., where one  
> can't
> accomplish the same thing in an LP-style language in a simpler way?)

See my comment above - more than are drempt of in your philosophy.

Ian


>
>
> 	--michael
>
>
>
>> Regards,
>>
>> Ian
>>
>> [1]
>> http://www.cs.man.ac.uk/~horrocks/Publications/download/2005/ 
>> HPPH05.pdf
>>
>>
>> On 23 Jun 2005, at 14:28, Bijan Parsia wrote:
>>
>>>
>>> On Jun 23, 2005, at 3:06 AM, Jos de Bruijn wrote:
>>>
>>>> -----BEGIN PGP SIGNED MESSAGE-----
>>>> Hash: SHA1
>>>>
>>>> Bijan Parsia wrote:
>>> [snip]
>>>>> Key is the slipping in of "ground".
>>>>
>>>> No. For query answering only ground entailment is relevant.\
>>>
>>> Interesting presumption.
>>>
>>>>>> entailments for both
>>>>>> semantics are *equivalent* and thus the queries would return the
>>>>>> same
>>>>>> result.
>>>>>
>>>>>
>>>>> Of course, RDF entailment includes existential generalization, so
>>>>> that's not quite right. There seems to be more work that you need  
>>>>> to
>>>>> do to get what you wanted (e.g., you need to look at the semantics
>>>>> of the query language; is the query "not" classical? how would that
>>>>> classical not interact with the LP semantics?)
>>>>
>>>> The RDF language contains existentials and I'm not claiming that  
>>>> this
>>>> can be done by a rule language.
>>>> We are talking about Horn Logic and Horn Logic does not have
>>>> existentials! I never claimed this!
>>>> You claimed
>>>
>>> Please point to where I claimed this.
>>>
>>>>  that a Horn formula under FOL semantics has other ground
>>>> entailments than a Horn formula under LP semantics and this is  
>>>> simply
>>>> not true.
>>>> I think this can conclude our discussion on this topic.
>>> [snip]
>>> Oh, I *quite* agree. But perhaps not for the same reasons.
>>>
>>> Cheers,
>>> Bijan.
>>>
>>>
>>
>>
>>
>
>

Received on Wednesday, 29 June 2005 21:59:08 UTC