- From: Stefan Decker <stefan@ISI.EDU>
- Date: Tue, 18 Nov 2003 01:46:16 +0000
- To: Jim Hendler <hendler@cs.umd.edu>, Benjamin Grosof <bgrosof@mit.edu>, Sandro Hawke <sandro@w3.org>
- Cc: adrianw@snet.net, www-rdf-rules@w3.org, phayes@ihmc.us
Please apologize my ignorance - what is hard about doing closed world reasoning on a giving RDF graph? Best, Stefan At 01:38 AM 11/18/2003, Jim Hendler wrote: >Ben - I think you miss my point - I didn't say figuring out a way to do >NAF would be a bad thing, I said it would be a very HARD thing, and one >for which there is no current de facto solution -- WOWG looked for a way >to do this, and realized we would not be able to do it -- I don't see why >the rules group would expect success unless they could start from an >existing solution -- and I've seen no proposal with a solution that seems >workable. If it's going to be part of the charter, then I would want to >see at least 1 workable solution before the WG starts... > -JH >p.s. WOWG's objective, which we didn't achieve, is mentioned in our >requirements [1] > >O3. Ability to state closed worlds Due to the size and rate of change >on the Web, the closed-world assumption (which states that anything that >cannot be inferred is assumed to be false) is inappropriate. However, >there are many situations where closed-world information would be useful. >Therefore, the language must be able to state that a given ontology can be >regarded as complete. This would then sanction additional inferences to be >drawn from that ontology. The precise semantics of such a statement (and >the corresponding set of inferences) remains to be defined, but examples >might include assuming complete property information about individuals, >assuming completeness of class-membership, and assuming exhaustiveness of >subclasses. Motivation: Shared ontologies, Inconsistency detection > >[1] http://www.w3.org/TR/webont-req/#section-objectives > > > >At 12:13 -0500 11/15/03, Benjamin Grosof wrote: > ><x-flowed>Hi Jim and all, > > > >At 03:18 PM 11/14/2003 -0500, Jim Hendler wrote: > >>Ben- > >> I agree w/Sandro - NAF requires identifying a set of facts it works over > >> (the domain) - but RDF graphs, but their very nature are open -- so what > >> sound easy suddenly becomes very hard. We attcked this problem in WebOnt > >> (see our reqs document and issues lists - sorry, I'm on slow connection > >> don't have the URIs, but they are one link from > >> http://www.w3.org/2001/sw/WebOnt) - we wanted a way to have a local > >> unique names assumption - but couldn't solve the problem -- I bet the > >> local domain naming is at least as hard, probably harder > > > >Would you please send me specific links when you can? I looked at the OWL > >requirements and issues list documents and I couldn't easily figure out > >which parts of them you were referring to. > > > >> here's an example, tell me whaty you would do > >> > >>You say > >> Rule1 - if person(shoesize) != large then A > >> Rule2 - if person(shirtsize) != large then B > >> RULES-CLOSED-OVER http://www.foo.bar/document1.rdf > >> > >>and that seems fine, but document1 includes > >> :Joe owl:class :person. > >> :Joe shoesize :large. > >> :Joe nickname "the gorilla". > >> :person rdf:type foo:human. > >> > >>now, foo is a namespace document which contains a bunch of facts about > humans. > >>It is clear that A is false, because the document you're closed over says > >>his shoesize is large > >>But what about B being true? We see that this document doesn't include > >>that his shirtsize isn large, but what is on foo:? Maybe it says anyone > >>with the nickname "the gorilla" where's a large shirt, maybe it refers to > >>another document, ad infinitum. > >> So when there is a web of graphs refering to terms in other graphs, etc > >> - how do you know where things stop? (see www-sw-meaning for a lot more > >> dicussion of this issue!) > >> this is also only one simple manifestation of this problem -- when you > >> talk about documents that are changing, scraped, etc. (all of which come > >> up on the web) it gets even uglier > >> > >> Sandro put it well - it's not that we cannot do NAF, it's that designing > >> the mechanism for definining the bounds of a graph on the web is still an > >> unsolved problem -- > > > >Thanks for the example, it helps. > >I think you've put your finger right on the nub of the problem. > >I was indeed presuming that there is a mechanism to define the bounds of > >the knowledge base / graph, i.e., to well-define the set of premises. > > > >> if the rules group has to solve it to make progess, that is risky > >> business.... > > > >I think the Semantic Web needs to solve it in an initial fashion, and quite > >soon. There's a tremendous overambitiousness in thinking that this is > >*not* critical path. It's not so hard to do, either -- in the following > >sense. Programming languages "solved" it long ago with mechanisms that > >check transitively for inclusion (such as the "make" facility in C). > >The obvious approach is to just use that type of idea for the Semantic > >Web. Thus if the transitive closure of the "import" chains cannot be > >determined and meet the usual criteria of well-definedness then there is a > >KB scope violation of a "system-ish" nature. This will force people to > >define more carefully exactly which portions of other KB's that they are > >importing -- including via more contentful module mechanisms within KB's -- > >and to do integrity checking on transitive closures of inclusion both > >initially when KB's are developed and periodically/dynamically as KB's are > >maintained/updated. > > > > I know that some don't like the idea of having to do this. I think the > >alternative of not being allowed to define such scoping is, however, > >extremely undesirable. The idea of "all RDF anywhere on the web" as > >something I would want to always *have to* use as my KB's scope is a > >complete non-starter practically -- consider issues of data/knowledge > >quality alone! (I'm tempted to say it's ridiculous. People talk about > >"trust" on the Semantic Web. The most basic mechanism for trust is simply > >to know what set of premises the inferences were drawn from. We'll be > >laughed out of town in most practical IT settings if we don't have a good > >story about this aspect of things.) > > > >If we take the approach I'm suggesting (and others have suggested it too) > >then we don't have to get fancy about deep philosophy and unplumbed > >territory of "social meaning", or wait for more research on "trust", to > >just get going on doing over the Web the kind of KR that has been proved in > >useful in decades of practical applications (and for a number of years in > >multi-agent systems). We can then proceed incrementally/evolutionarily > >over time, as we develop further use cases and techniques, to open things > >up by having more implicit and relaxed mechanisms for importing / scoping > >the KB's/graphs. We should start with what we know works, in short, and > >then work to improve upon it in the direction of reducing the burden of > >defining inclusion/import scoping. As a practical matter, if there is a KB > >scope violation cf. above, then that doesn't mean we can't/won't do > >inferencing, depending on the purpose and kind of inferencing -- some kind > >of inferencing may be useful even when there is a violation. > > > >If we do it that way, we can have/do nonmon/NAF on the Semantic Web > >essentially today, and develop additional techniques later for making the > >scoping more flexible and convenient. > > > > > >> -JH > >>p.s. Note that the OWL group rjected the solution that we could use the > >>imports closure and define everything else as not included, because that > >>would limit you to only those things defined in the DL profile, not all > >>OWL and all RDF documents > > > >I'm confused by this. "All OWL and all RDF documents" is way too big -- > >see above my comment about "all RDF on the Web". When you say "DL profile" > >I presume you mean the set of OWL imports statements. What's the point of > >an imports mechanism in OWL if everything else is included? Perhaps I'm > >not understanding what you're saying. > > > >In any event, the way to go is to define (a given KB as) importing of RDF > >as well as OWL (and soon, more generally, semantic web rules knowledge base > >modules as well), in the imports profile, and stick to the transitive > >closure for most purposes. Does that require extending the current imports > >mechanism of OWL, e.g., to define a boundaried RDF graph as imported? > > > >>-- the rules language would have to face that same issue, but also deal > >>with all things findable by Xquery ... yow! > > > >I don't see what XQuery has to do with it (at least not directly), if we're > >talking RDF stuff. XQuery is certainly related to Semantic Web Rules > >(indeed, I was one of the first to press this point to the W3C team; back > >in March 2001 I presented to them about it), but I don't see that Rules > >"have to... deal with all things findable by XQuery". More pertinent to > >the main topic here is that XQuery deals quite ambitiously with very large > >scale databases and as I understand it (from early versions I looked at) > >has a well-defined boundary of what is queried over. That's thus probably > >further evidence towards the usefulness of my scoping suggestion about > >imports closure. > > > >Benjamin > > > >>-- > >>Professor James > Hendler http://www.cs.umd.edu/users/hendler > >>Director, Semantic Web and Agent Technologies 301-405-2696 > >>Maryland Information and Network Dynamics Lab. 301-405-6707 (Fax) > >>Univ of Maryland, College Park, MD 20742 240-277-3388 (Cell) > > > >_________________________________________________________________________ > _______________________ > >Prof. Benjamin Grosof > >Web Technologies for E-Commerce, Business Policies, E-Contracting, Rules, > >XML, Agents, Semantic Web Services > >MIT Sloan School of Management, Information Technology group > >http://ebusiness.mit.edu/bgrosof or http://www.mit.edu/~bgrosof > > > ></x-flowed> > > >-- >Professor James Hendler http://www.cs.umd.edu/users/hendler >Director, Semantic Web and Agent Technologies 301-405-2696 >Maryland Information and Network Dynamics Lab. 301-405-6707 (Fax) >Univ of Maryland, College Park, MD 20742 240-277-3388 (Cell) > -- http://www.isi.edu/~stefan
Received on Monday, 17 November 2003 20:46:29 UTC