Re: Existentially Quantified Variables (bNodes) in RDF-like-languages

Uh oh.  You called URIs "unique" names when I think they are
"unambiguous".  By my understanding of any system we're trying to
build, one thing can have any numbers of URIs, but two things with the
same URI are the same thing.  Do you think I am wrong?

I think all your points come down to this issue, so I'll hold off on a
point-by-point reply until this is settled.

    -- sandro



> >This is an old issue.  Pat Hayes explained very nicely a year ago [1] that
> >RDF M&S has always had existentially-quantified variables (in the subject
> >role), and to take them out would be a change.  Many others (including mysel
> f
> >at the time) supported this position, including the RDF Core WG.  Some
> >agreed that's what M&S said, others simply thought existentials would
> >be good to have.
> >
> >But I want to revisit the question more fundamentally.  What good are
> >existentially quantified variables in a language like RDF? [2] Why would we
> >possibly want them?  I no longer know a good reason.
> >
> >Here are four reasons I have heard for them, and my counter-arguments.
> >
> >Reason 1:  Because people often express themselves without giving URIs for
> >            things.  Using pronouns like "something" feels like a very natur
> al
> >            and comfortable way to communicate.
> >
> >	   Response: While we may not explicitely have names for things, the
> >	   ability of natural language to "go meta" means you can always talk
> >	   about "the thing you meant when you said '...'".   So we can still
> >	   identify the denotations of terms used in conversations, even if no
> >	   explicit identification was given at the time.
> 
> Well, yes and no. What you are talking about here isnt really going 
> META, its more like referring back to something in a CONVERSATIONAL 
> context, what linguists call the 'common ground'. And that is the 
> point: in order to do that, you have to have a common ground to refer 
> to. Imagine someone who just walks up to you and says 'remember that 
> thing you were talking about...' and it turns out that he is 
> referring to a conversation you and he had several years ago. (This 
> has actually happened to me, by the way; the person in question hadnt 
> seen me for almost a decade, and he phoned one day and just carried 
> on the conversation.) The point of the example being that in order 
> for this to work, you both need to have some way to refer to the 
> context of the conversation in which the name was used; and 
> remembering enough stuff in order to be able to make sense of such a 
> back reference is a major computational cost. So major, in fact, that 
> it isnt reasonable to expect that any agent does maintain all such 
> contexts.
> 
> Insisting on having globally unique URIs for every occurrence of 
> every local variable amounts to saying that the Web should somehow be 
> able to remember every use of every blank node in every context, for 
> ever. Maybe it would be logically nice if that could be done, but it 
> clearly can't be done, and it just won't happen.
> 
> >
> >	   Alternatively, who cares.  This is a formal system, and 
> >any appeals to
> >	   human qualities are likely to be misleading anyway.
> >
> >Reason 2:  We were getting a lot of noisy URIs in our knowledge bases.
> >
> >	   Response: This isn't really a model-theory issue.
> 
> It is a real issue, however, and I think it is central. The point is 
> not just the syntactic clutter, but the inference patterns that these 
> names will block. If I copy some RDF then I do not have a licence to 
> rename any of the urirefs in it. But I *should* have such a licence 
> for the blank-node urirefs that you want to insert, because those 
> urirefs don't actually mean anything; they are just arbitrary labels. 
> .
> 
> >  A
> >	   best-practices approach might be to suggest that urn:uuid's be
> >	   used for any identifiers which are both not meant for humans and
> >	   not tied to any on-the-web knowledge base.  User interfaces
> >	   shouldn't be showing URIs anyway, but they could hide these even
> >	   more.
> >
> >	   Another approach is to have meta-data, with triples saying that
> >	   the URI "urn:uuid:57c259ca-47ed-11d6-8002-0050ba4812a6" is not
> >	   meant for humans, or something.   (Note that such an assertion
> >	   must be about the above string, not the object denoted by that
> >	   string.   That object may be denoted by many URIs, some of which
> >	   are prettier than others.)
> >
> >Reason 3:  Because we can use existential variables for querying RDF.
> >            (Pat suggests this in [1].  It was my reason a year ago.)
> >
> >            Response: I've seen a lot of RDF query languages now, and this
> >            approach is not being used.
> 
> It is in the proposed DAML query language.
> 
> >  In my own query language designs,
> >            I've found them unsatisfactory.
> >
> >Reason 4:  One should be able to commit to a statement like "Someone on E-Ba
> y
> >            is selling a signed copy of 'Weaving the Web'."  That statement 
> is
> >            different from "Someone on E-Bay is selling the signed 
> >copy of 'Weaving
> >            the Web' which has the URI http://example.com/copy7".  The secon
> d
> >            one entails the first, of course.
> >
> >	   Response: Why do you want to make the first statement?  What bad
> >	   thing will happen if you call it
> >	   <urn:uuid:57c259ca-47ed-11d6-8002-0050ba4812a6> ?
> 
> A better question would be, what bad thing would happen if you said 
> <urn:uuid:5-0850-123985=0-129ut0jfopqi7>  is selling a signed 
> copy..., and then some inference engine tried to figure out 
> <urn:uuid:5-0850-123985=0-129ut0jfopqi7>'s SS number, knowing that he 
> is a person living in the USA.
> 
> But a direct answer is that one bad thing would be that any inference 
> that did not use that particular name would be blocked.
> 
> >
> >	   Are you worried that someone else will say something about
> >	   <urn:uuid:57c259ca-47ed-11d6-8002-0050ba4812a6>?
> 
> No, the point is more the other way round: if they want to say 
> anything about the thing that is for sale, they are obliged to use 
> that particular uriref. It is unique, remember. Also, if you 
> generated it, then you are responsible for remembering what it is 
> supposed to denote. For ever.
> 
> >  Like "it's
> >	   missing page 7?"  But they could do that anyway (with existentials)
> >	   by using various more-expressive ontologies, perhaps talking about
> >	   all the books on E-Bay, or talking about the terms in your
> >	   original statement.
> >
> >	   When you used a URI for E-Bay, were you committing to any
> >            facts?
> 
> That the thing in question has a certain name, yes. Suppose another 
> engine, given all the facts about 
> <urn:uuid:57c259ca-47ed-11d6-8002-0050ba4812a6>, wants to use some 
> other name for it and assert those facts using that other name. Is 
> that a logical crime?
> 
> >  Which ones?     I think that's the real issue here. 
> >
> >	   I think a URI enables a client to learn more about a thing,
> >	   but the use of a URI in one document cannot be taken to
> >	   mean that someone asserting the document believes any facts
> >	   outside the document.  
> >
> >	   ... but then how do you know what was meant?  If I write
> >	   you a check for $100, how do you know if I meant $US or
> >	   $Canadian?  The fuzzy vision of the Semantic Web involves
> >	   the meaning of URIs being established by some kind of
> >	   consensus reality, I think.  But how does one commit to a
> >	   statement in such fuzzy terms?  
> >
> >	   Anyway, maybe existentials will be useful as being free
> >	   from consensus reality issues, but I don't see how.  I want
> >	   to see the argument you present to a judge when you're
> >	   suing someone for breech-of-RDF-assertion.  Maybe it comes
> >	   down to the natural-language comments in the ontology.  Do
> >	   you need to explicitely commit to an ontology, or just use
> >	   terms from it.  What if two ontologies exist with the same
> >	   URIs?  Which one are you committing to?  I think the only
> >	   solid ground is to explicitely import facts (including
> >	   natural-language documentation) from some URI, giving a
> >	   secure hash of the contents if you're worried it might
> >	   change.   (but that's not the way RDF is being used today...)
> >
> >The reason to leave them out is that they complicate things.  
> >
> >Its worth noting, too, that if you really want to make an existential
> >statement, you can do it in some higher-layer logic language.
> 
> That is quite unclear. It depends on the higher layer and how it is 
> connected to RDF. If there are no anonymous nodes in RDF-lite, how 
> does it handle containers? If it has no containers, how does it 
> encode higher-level syntax?
> 
> Pat Hayes
> -- 
> ---------------------------------------------------------------------
> IHMC					(850)434 8903   home
> 40 South Alcaniz St.			(850)202 4416   office
> Pensacola,  FL 32501			(850)202 4440   fax
> phayes@ai.uwf.edu 
> http://www.coginst.uwf.edu/~phayes

Received on Friday, 5 April 2002 11:00:28 UTC