RE: Punning discussion

Hi Bijan!

Bijan Parsia wrote:

>On 24 Jan 2008, at 14:41, Michael Schneider wrote:
>
>> Bijan Parsia wrote:
>>
>>> On 24 Jan 2008, at 10:48, Michael Schneider wrote:
>>> [snip]
>>>>> From: Alan Ruttenberg
>>>>> Sent: Tuesday, January 22, 2008 9:21 AM
>>>>
>>>>> 3) [...] How does punning effect OWL Full?
>>>>
>>>> Punning is a DL-only feature,
>>>
>>> This doesn't seem right. If you take the equality free fragment of
>>> OWL Full, you cannot distinguish between punning and hilog 
>semantics.
>>> RDF, the language, doesn't have equality, so it is compatible with
>>> punning.
>>
>> The aim is to provide an RDF compatible semantics, which has  
>> additional
>> semantics for all OWL-1.1 language features.
>
>What this means is open to discussion.

IMO, discussion works best on existing drafts, as I have learnt from those good editor's drafts from which the WG originally started. So that's what I am going to help creating for OWL-1.1-Full, at least as good as I can within my restricted abilities and the time remaining. Allow me a few weeks for coming up with something, you actually had several years to produce the other drafts.

>> And the WG charter wants
>> OWL-1.1-Full to maintain backwards compatibility with OWL-1.0-Full.
>
>And would like to extend the compatibility, this might mean making  
>slight changes to both. In either case, these are SHOULDs not MUSTs

(Please...)

Ok, maybe we should(!) try to settle on the meaning of the word "SHOULD" here. So my question to you or to the many people around, which are more experienced in W3C WG activities than me (i.e. everyone else): Is there any documented meaning of the word "SHOULD" in the W3C? 

Otherwise, I will have to come up with my own interpretation, and you may decide whether it is acceptable for you or not.

>and it seems reasonable to see whether we can do better overall.
>[snip]
>> And to answer your statement above: There is, of course, an  
>> 'owl:sameAs'
>> construct both in the 1.1-RDF mapping, as in the 1.0-Full spec, and  
>> it has a
>> semantics in the latter. So I don't see how talking about "the  
>> equality free
>> fragment of OWL-Full" does help us in this discussion.
>
>Well, you made a claim that somehow punning and OWL Full semantics  
>were so radically different as to be incompatible. I pointed out a  
>clear case where they coincide. 

Can you please elaborate a bit more on this? I can only guess what you mean here. Maybe you want to say something like that RDFS would produce the same entailments under punning semantics, too? Can you please give me some more hints? Pointers?

>I also, several times, have pointed  
>out how the "mapping to one thing" style of RDF could be preserved  
>and that, since RDF says nothing about equality, 
>how we handle equality is more or less up to us.

... /would/ perhaps be up to us, if there wasn't OWL-1.0-Full around, which already has equality defined. But now we have to (ok: "should") maintain backwards compatibility.

>>>> and cannot affect the definition of Full, as
>>>> long as Full is understood to be an RDF compatible language.
>>>
>>> The semantics specs something in a style that is *suggestive* of how
>>> to extend it to OWL, but that's a different thing. It doesn't say
>>> *how* to extend the semantics. It seems as legitimate to add
>>> parameterized extension relations as anything else.
>>
>> There is precedence given by the OWL-1.0-Full spec.
>
>That's a *different* argument, i.e., it's not that OWL 1.0 Full is  
>the only way to have an "RDF compatible model theory". 

Perhaps... Well, that's speculative, since now that OWL-1.0-Full is there as a precedence, we (I know, repetition is boring) *should* maintain backwards compatibility.

>You argued
>that we had no jurisdiction to change the RDF semantics. Let's say I  
>agree.

I hope so, since otherwise I would be worried about you stealthily messing around with the original RDF spec documents at the W3C web server. ;-)

> But then we *do* have jurisdiction to change OWL Full (as we  
>are going to update it after all). 

As far as it remains RDF compatible. And as far as it remains backwards compatible to OWL-1.0-Full. Yes, I think there still remains quite some freedom for creativity within the given constraints. :)

>Your argument against punning 
>as the way we handle OWL Full is that its incompatible with RDF  
>semantics. 

Is this surprising? No one asks for RDF compatibility of OWL-DL. Neither was OWL-1.0-DL RDF compatible, nor does OWL-1.1-DL need to be. It is actually the characteristic feature of OWL-Full to be an (or better: "the") RDF compatible variant of OWL.

>I've shown that's not true.

I'm not sure if I understand. Do you say that you have shown (or wanted to show) that it is _not_ true that OWL-1.1-DL+Punning is _not_ compatible with RDF? 

>In this way I also have shown that punning is compatible with OWL  
>Full with hilog semantics. 
>(I.e., one can do the "One thing" mapping,  
>have the parameterized IEXTs, then, in OWL Full, add an additional  
>semantic condition to force extension merging).

Again, I don't understand what you exactly mean here. Can you please elaborate? The main problem I have in understanding is probably, what you exactly mean by "compatible" here, and how you define the "parameterized IEXTs". 

>[snip]
>>> From the same paper I can see that it really seems to be possible  
>>> to have
>> real metamodelling
>
>I don't find that hilog metamodelling to be more real than punning 
>or, for that matter, less real that "true" second order logic or  
>stratified approaches.
>
>> (definition 6), at least in SHOIQ(D), without falling
>> into undecidability.
>
>You need to impose a unique name assumption on role names.

That's bad, admittedly! :-(

>> But I haven't read the details yet, so you or Boris or
>> anyone else may FIXME if I'm wrong. Nevertheless, it seems that it  
>> has not
>> been decided to use this "HiLog" semantics in 1.1-DL, for whatever  
>> reason,
>> but to use the weaker "contextual" (aka "punning") semantics instead.
>
>There is no implementation experience (Boris and I have discussed  
>some ideas but they could have a fairly large performance effect).

Ok.

>It's pretty clearly not what people expect. 

This is just a claim. You need to show me evidence, since it really sounds unbelievable to me. I can just claim the opposite (and I'll do so, see below!).

>Punning has a long history in Pellet 
>so we have both user and implementor experience.

Pellet has punning support, granted. Pellet has (many) users, granted. But does this mean that punning support in Pellet has users? I'm curious, can you please provide some evidence for actual punning users?

>It's also forward compatible as in you can add hilog semantics ontop  
>of punning (as exhaustively discussed). So there's a pretty clear  
>path forward if it turns out that people want and can use the hilog  
>semantics.

This is certainly technically true. (And really seems to be in the spirit of "incrementally" upgrading OWL. ;-))

>> If you now lift your hand and honestly tell me that OWL-1.1-DL  
>> allows this
>> so called "HiLog semantics", too
>
>What? Why would I say this? it's clearly false. And with excellent  
>reason.
>
>> (and, perhaps, that its left to the
>> implementors of OWL-1.1 reasoners to decide which of these  
>> semantics to
>> support), then I promise to sit myself down and carefully check the
>> technical details closely, and it may well be that I will never  
>> stand up
>> again and say a single evil word against "punning". :)
>
>Sorry, I have never said, nor would I ever suggest, that punning  
>includes hilog semantics. It's exactly as in boris's paper.

I haven't claimed that you have said this. Hm, perhaps it's because the term "if" in the above stated sentence may be read with two different meanings? I suppose, I should learn to clearly distinguish between different meanings by using different terms to denote them (here: "should").

>But that they are not the same doesn't make them incompatible. Given  
>that there is no implementation of Hilog semantics, I conclude at  
>this stage of the game that it's not used. It causes a number of  
>problems and issues.
>
>But if you want to include it in an OWL Full, fine. 
>But you've  
>already talked about giving up syntax reflection (in the precise  
>style of OWL Full) 

[Not in this list. And the discussion then now seems a bit obsolete, I suppose.] 

>so once you've gone there I think the burden of  
>proof shifts a bit.
>
>By the way, at just about every opportunity for over 3 years I've  
>explained the difference between punning, hilog, and syntax  
>reflection to as many users as I can. I spent on session discussing  
>it at the first OWLED. I keep trying to get panels etc. etc. My  
>experience is that the hilog (and syntax reflection), at least in  
>full generality, is not anything people know about or when informed  
>that they can see any use for. Or would be particularly good at  
>working with. 

Ok, granted for syntax reflection. I wouldn't bother people with this (well, except perhaps a few very good friends of mine :)). But what is actually the problem with metamodelling? I still haven't looked at the gory details of HiLog semantics, but from my general understanding I don't see the slightest problem where one could fail to understand (what I call) "real" metamodelling - even in "full" generality. Perhaps, we are talking about different things here, so I am going to present here how I would explain it to people:

Imagine you have two relations, named 'r' and 's', and you know that r = s. What can you say about them? First, in the way used in the equality expression, r and s are used as individuals. But since r and s are relations, you can conclude that if a relationship r(x,y) between two individuals x and y is true, than also the relationship s(x,y) is true, and vice versa. So r and s have an equal relation extention. But you can also regard r and s as classes: If z is an element of r (z is actually a tuple, but this is not important here), then z is an element of s, and vice versa. So r and s have an equal class extention. And, of course, you can have the relations r and s be themself elements of some class C, or they may stand in some relationship p(r,s), where p is some other relation. So r and s are individuals under this point of view, again.

Now, was this in any form hard to understand? Or hard to believe? Or is it supposed to be hard to work with? Or does it look like something, what might perhaps be right, but seems completely useless? (Ok, the latter perhaps, since its mathematics, after all. ;-))

>I worry a little about putting hilog into owl full (again) 
>because when we *do* get a user base who understands and  
>wants to use "implication heavy" metamodelling, there's no clear  
>evidence that hilog (as opposed to stratified, for example) is what  
>they'll want.

So /here/ you are suddenly asking for evidence yourself. Well, should you like my explanation above, you may freely use it everytime and everywhere to probe everyone's understanding of metamodelling! :)

>Standardization's primary goal, IMHO, is to support interoperability.
>I don't mind adding new features if we're likely to get multiple  
>implementations (i.e., raise the floor), but this just doesn't seem  
>to be the case here.
>
>Cheers,
>Bijan.

Cheers,
Michael

--
Dipl.-Inform. Michael Schneider
FZI Forschungszentrum Informatik Karlsruhe
Abtl. Information Process Engineering (IPE)
Tel  : +49-721-9654-726
Fax  : +49-721-9654-727
Email: Michael.Schneider@fzi.de
Web  : http://www.fzi.de/ipe/eng/mitarbeiter.php?id=555

FZI Forschungszentrum Informatik an der Universität Karlsruhe
Haid-und-Neu-Str. 10-14, D-76131 Karlsruhe
Tel.: +49-721-9654-0, Fax: +49-721-9654-959
Stiftung des bürgerlichen Rechts
Az: 14-0563.1 Regierungspräsidium Karlsruhe
Vorstand: Rüdiger Dillmann, Michael Flor, Jivka Ovtcharova, Rudi Studer
Vorsitzender des Kuratoriums: Ministerialdirigent Günther Leßnerkraus