- From: Michael Schneider <schneid@fzi.de>
- Date: Wed, 25 May 2011 13:47:27 +0200
- To: <markus.kroetzsch@comlab.ox.ac.uk>
- CC: <antoine.zimmermann@insa-lyon.fr>, W3C SWIG Mailing-List <semantic-web@w3.org>
Hi! On Thu, 19 May 2011 Markus Krötzsch wrote: >> What Markus says here I guess is that, in spite of the limitations of >> the punning mechanism, a full-fledged OWL 2 DL reasoners will likely >> infer more things than *currently existing* incomplete OWL Full >> reasoners. > > Right. Not right! See my yesterday's mail: <http://lists.w3.org/Archives/Public/semantic-web/2011May/0189.html> Even many of the most "light-weight" RDF entailment-rule reasoners will give you at least some of the metamodeling-related results that you cannot get from any conformant OWL 2 DL reasoner, provided that these rule reasoners support rdf:type together with owl:sameAs-based substitution of nodes in a graph. Just play around with an arbitrary RDF triple store that provides for some basic inferencing. And the more expressive such "RDF reasoners" get, the further they go beyond OWL 2 DL concerning metamodeling-based results, while still keeping massively incomplete w.r.t. OWL 2 Full: Jena's OWL reasoners, OWLIM... there are many around. > We know that there cannot be a tool that computes all > consequences of OWL with "proper" meta modelling, No, that's *not* what we know! Taking the term "consequence" to mean the same as "logical entailment", and taking "OWL with proper meta modeling" to mean "OWL Full", then all we know is that there cannot be a tool that computes all consequences AND all *non*-consequences of OWL Full. What you claim here is basically that OWL Full isn't even *semi*-decidable, which has never been proven by anyone. In fact, the whole semantics specification of OWL (1/2) Full, i.e. the set of model-theoretic semantic conditions that build the semantics of OWL Full (the OWL 2 RDF-Based Semantics), is defined as a set of standard first-order formulae: OWL 2 Full is, essentially, defined by first-order theory! So it should be clear (at least to those people knowing about logics) that OWL (1/2) Full *is* semi-decidable. And, of course, this means that there *can* be tools that compute all consequences of OWL Full. There will just be no complete tools for computing all *non*-entailments. > and we also > know that some forms of meta modelling can even lead to > intricate inconsistencies that make the whole ontology > language paradoxical (PF Patel-Schneider's paper "Building > the Semantic Web Tower from RDF Straw" alludes to this > issue). We *knew* this for OWL 1 Full. The results in the cited paper fundamentally depend on so called "comprehension conditions", which were part of the semantics of OWL 1 Full, but are not normative in OWL 2 Full anymore. A different approach called "Balancing" has taken the place of the comprehension conditions, and so the issue has gone and the paper (at least the argument you refer to) is moot. > So it seems that a tool that obtains all consequences > of plain OWL constructs, and that can still handle some > meta modelling is not such a bad choice, even if it is > called "OWL DL reasoner" ;-) To me it seems, that a tool that obtains all consequences of plain OWL constructs, and that can still handle *all* meta modelling, while occasionally not coming back when processing non-consequences, might be even a better choice. And that's what we can actually expect from an "OWL Full reasoner" (when using "Balancing"). Btw, being semi-decidable (or recursively enumerable) is sufficient to produce complete answers under the SPARQL 1.1 RDF-Based entailment regime (because all we need is complete enumeration of answers), and this is pretty much what I consider to be sufficient to me when I am about doing reasoning on the Web. Complete decision (where the stress is on *non*-entailment detection) is certainly a plus in some scenarios, but my main interest in practice is on getting inferences, not so much on learning what is not an inference (except when I am doing analysis work). But even in scenarios where the detection of non-entailments or consistent ontologies is of real relevance, keep in mind that undecidability does not mean that I will always get no result; it just means that, theoretically, there *exists* some input for which my reasoner won't come back. Whether this has any relevance for practical reasoning is completely unclear from the pure fact that a language is undecidable. And in practice, there will always be a maximal a-priory reasoning time to be granted to the reasoner, so whether there is no result from undecidability or for whatever reason is pretty irrelevant (and there are tons of other reasons for a reasoner to not come back in time). In any case, one has to check *experimentally* for one's specific application if it works or not, regardless whether one uses decidable or undecidable reasoning. And I can tell you that successful decision under undecidable entailment regimes is not so uncommon... I know, because I *have* checked! Cheers, Michael PS: I just found that my spell-checker did not know the term "undecidability". That lucky one! -- Dipl.-Inform. Michael Schneider Research Scientist, Information Process Engineering (IPE) Tel : +49-721-9654-726 Fax : +49-721-9654-727 Email: michael.schneider@fzi.de WWW : http://www.fzi.de/michael.schneider ============================================================================== 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 Stiftung Az: 14-0563.1 Regierungspräsidium Karlsruhe Vorstand: Dipl. Wi.-Ing. Michael Flor, Prof. Dr. rer. nat. Ralf Reussner, Prof. Dr. rer. nat. Dr. h.c. Wolffried Stucky, Prof. Dr. rer. nat. Rudi Studer Vorsitzender des Kuratoriums: Ministerialdirigent Günther Leßnerkraus ==============================================================================
Received on Wednesday, 25 May 2011 11:48:00 UTC