Re: To embed or combine

Dave,
I get your point about the entailment regimes.
But it seems to me that different entailment regimes require different axioms
to be added to the standard embedding. So, these can be handled this way,
and this can be done in a clean and modular way.

It might seem strange that somebody like me is arguing against including a
model-theoretic semantics. But in this case it really looks redundant to
me. All I am asking is that people in the group look at that stuff and
decide for themselves whether the model theory specified in the RDF/RIF
document appears redundant to them as well.


	--michael  

> On reflection I think we have to slice this slightly differently.
> 
> There's the question of how RIF rule sets can simply access RDF data. I 
> believe that requires parts of the "Embedding triples and Graphs" 
> section to be normative (see below).
> 
> Then there's the question of the various RDF(S) entailment regimes and 
> how those interact with RIF. I think those are specified better by means 
> of the model theoretic semantics (see below).
> 
> I also think we need a separate informative document (or document part) 
> which offers a "Guide to using RIF with the semantic web" that is more 
> accessible to implementers and users and which opens up the options of 
> non-normative subset semantics.
> 
> 
> ** Accessing RDF data
> 
> We have a UCR requirement that RIF rules should be able to "cover RDF 
> triples as data where compatible with phase 1 semantics". To me this 
> means it should be possible write RIF rules which can access RDF 
> statements.
> 
> This requires the syntax section (how literals and URIs are mapped) and 
> the tr and tr\sub(s) mappings from "Embedding triples and Graphs" to be 
> normative.
> 
> Further, as I've argued several times, in practice RDF rule languages 
> include builtins like "isBlankNode". Supporting those requires a 
> modification to the tr mapping to distinguish the skolem constants as 
> referenced by the blue text in Jos' document.
> 
> ** Entailment regimes
> 
> Since RDF is not simply data but also comes with a set of entailment 
> regimes we also have to specify how RIF rule sets that access RDF data 
> interact with those regimes.
> 
> [The RIF processor will also need to know what the required RDF 
> entailment regime is - hence the Data Sets section in Arch.]
> 
> I think the interaction of the entailment regimes is best specified 
> using a model theoretic approach as Jos has done. I say that for the 
> following reasons:
> 
> o In the RDF specification the rule sets for RDF and RDFS entailment are 
> purely informative, it is preferable for the RIF specification to 
> reference the normative parts of the RDF specification.
> 
> o We agreed in RIF that dialect semantics should be specified model 
> theoretically where possible, in preference to proof-theoretic or 
> operational semantics. That seems consistent with preferring a model 
> theoretic formulation of how the RDF and RIF entailment regimes interact.
> 
> o If we made the RIF rule sets for RDF and RDFS entailment normative we 
> would be directing implementers to use this approach. Whereas in 
> practice RDF implementations will already have solutions to especially 
> the RDF entailment and should be free to use that and connect to a RIF 
> processor in "black box" style if desired.
> 
> o The model theoretic approach seems more extensible towards defining 
> the interaction with OWL (which we also have to do) than one based on 
> translation rules.
> 
> At first I had been concerned that Jos' proposal forces the full 
> RDF/RDFS semantics on implementations whereas in practice many people 
> implement subsets of the entailments. However, when people want to 
> support subset semantics they can do this by specifying just simple 
> entailment and conveying the subset semantics by imported RIF rulesets, 
> such as that for rho-df. That practice is sufficiently important that I 
> think we should enable it by providing a rule import mechanism and 
> document it in a non-normative "Guide to using RIF with the semantic web".
> 
> We could simply duck the question of how RDF and RIF semantics interact 
> altogether. We could just specify the data access embedding and give 
> neither the model theoretic nor the rule-based-embedding normative 
> status. I don't think that would be ideal but would probably be 
> acceptable to me.
> 
> So my conclusion is to restructure the document slightly:
>    o RDF graph embedding (normative) combining current syntax section 
> and definition of tr and tr\sub(s)
>    o Semantics of RIF rules combined with RDF entailment, model 
> theoretic, normative
>    o Embedding RDF semantics, informative
> 
> Note that putting the tr definition in the first section also clarifies 
> for implementers what is actually going on and may help to alleviate 
> some of Michael's concerns.
> 
> Dave
> -- 
> Hewlett-Packard Limited
> Registered Office: Cain Road, Bracknell, Berks RG12 1HN
> Registered No: 690597 England
> 
> Chris Welty wrote:
> > 
> > 
> > <chair>
> > The status of the discussion regarding Jos' RDF compatibility section 
> > appears to be mired in whether the normative semantics of RDF in RIF 
> > should be specified in the model theory through a "combination" of RIF 
> > and RDF semantics, or through an "embedding" of RDF semantics in RIF (as 
> > rules).  The two approaches have been shown by Jos to be equivalent.
> > 
> > At the moment I have not seen any technical arguments supporting one 
> > approach or the other.  Michael prefers the "embedding" on the basis that:
> > 
> > (1) the "combination" is more complicated than the "embedding" and thus 
> > more difficult to understand.
> > 
> > (2) it is not our job viz. our charter to specify a model theoretic 
> > approach to the RDF/RIF combination
> > 
> > Jos seems to prefer the "combination" and argues re: (1) that:
> > 
> > (3) it is no more difficult to understand the "combination" than the RIF 
> > model theory.
> > 
> > As chair, my own read of the charter does not provide any particular 
> > help on (2), I'm not quite sure what Michael is referring to there.  It 
> > is certainly our job to specify how RIF and RDF should be used together, 
> > and as chair I interpret this as meaning we should have a normative 
> > standard for that.
> > 
> > Thus, as suggested by Michael, it seems to me we are at a difference in 
> > preference only, and I see no alternative other than to call a vote.  It 
> > seems to me the vote is about which approaches to make *normative*:
> > 
> > 1) The model-theoretic "combination" of RIF and RDF is normative
> > 2) The "embedding" of RDF semantics as RIF rules is normative
> > 3) Both the "combination" and "embedding" are normative (What would that 
> > mean?)
> > </chair>
> > 
> > -Chris
> > 
> > 
> 
>

Received on Monday, 10 September 2007 18:59:04 UTC