Re: 'shared meaning' landgrab?

From: Dan Brickley <danbri@w3.org>
Subject: 'shared meaning' landgrab? 
Date: Fri, 23 Aug 2002 17:26:34 -0400 (EDT)

> 
> Hi Peter,
> 
> On Fri, 23 Aug 2002, Peter F. Patel-Schneider wrote:
> 
> > From: "Seth Russell" <seth@robustai.net>
> > Subject: Re: A Rough Guide to Notation3
> > Date: Fri, 23 Aug 2002 10:03:15 -0700
> >
> >
> > > Why?  Strangely enough  I happen to believe it is the current obsession with
> > > applying formal logic to language that is one of the main problems with the
> > > Semantic Web.  Imho, the Semantic Web is about communication and finding
> > > whatever we are looking for.  I have yet to see a single application of
> > > formal logic that has furthered that goal.
> 
> > Well, how are you going to ascribe a common meaning to the graphs you are
> > sending around?
> 
> This was the main point I wanted to engage with, and the reason for the
> change in Subject: line... I believe comments such as yours re 'common
> meaning' to be over-enthusiastic marketing for a technical approach which
> strangely enough *avoids* engaging with many of the challenges of 'common
> meaning'. 

First, I'm not marketing anything (at least not yet).  The marketing is on
the RDF side.  

> The proposed Model Theoretic formalisations of RDF, OWL etc only
> scratch the surface of this problem. It is imho simply incorrect to
> suggest that the RDF/OWL formalisations fix a 'common meaning' for such
> content. They are a powerful tool, but not only address limited aspects of
> the problem.

> A Model Theory for (Semantic)Web content shows us how to stick to
> truth-preserving operations on Web documents/messages with propositional
> content, ie. those XML documents that are candidates for being true or
> false (in some interpretation). But the MT approach is, by design, silent
> on the imho more interesting part of the puzzle. Namely, the problem of
> fixing the shared propositional content of these RDF/XML documents, such
> that scattered parties throughout the Web can use them to communicate. The
> MT specs alone can't fix the content of the RDF/XML, ie. determine the
> claims they make about the world.

Formal languages, by themselves go only part of the way, no matter which
formal method of assigning meaning to their constructs is used.  There
always has to be some ``connection to the real world'' to ground the formal
language.  However, without some formal meaning for a langauge, there is no
hope of using that language for building up machine understandable
documents in the Semantic Web.

Look at XML, for example.  XML is a perfectly good formal language.  It
even has a formal meaning, of sorts.  XML can thus be used to send around
machine understandable documents.  However, the formal meaning of XML is
poor at composing meaning of XML documents from their pieces, so programs
that understand the XML documents they pass around have to have much of
that understanding built in.   The goal of the Semantic Web, as I see it,
is to allow the building up of such understanding.

> >From http://www.w3.org/TR/rdf-mt/ "Since the exact nature of the things in
> the universe is irrelevant, it is convenient to use the nodes of the graph
> themselves as their own denotations."
> 
> MTs are happy with such trickery, since they only care that there are
> satisfying interpetations, and not whether the symbol->world mapping
> corresponds to the "common meaning" that people have in mind when encoding
> claims in RDF/XML.

Of course.  Formal meanings have to be built up from something, and it is
generally better to put as few constraints on this foundation as possible.
Somehow, the base layer of meaning has to be provided using other means.
The aim of many researchers in knowledge representation is to provide
languages and tools that make it possible to build up a shared meaning for
complex constructs, based on a meaning for some components.

> So it strikes me that you're making something of an unwarranted land-grab
> on the term 'shared meaning'. I agree that MTs may help us detect and
> repair miscommunications, but it seems too strong to claim that
> formalisations in the MT tradition are enough to fix a common 'shared
> meaning' for RDF/XML documents. 

Well, MTs (and formal meaning in general) does much more than help and
repair miscommunications.  They fix the meaning of the structure of the
language, so that it can be used to communicate `shared meaning'.  For
example, without an account that says that the meaning of an RDF graph is the
conjunction of the meaning of its triples, how can an RDF graph be used to
communicate any meaning at all?

> Other ingredients come into play here,
> notably the URI specifications. Unlike MTs, the URI spec does address this
> aspect of the shared meaning problem, since it is concerned with
> establishing machinery by which can write down names for things in an
> agreed syntax and with common (if tangled, poorly articulated) rules about
> the mapping between symbol and world.

I don't see how the URI spec does this at all.  RDF treats URIs as opaque
identifiers, after all, so, as far as RDF is concerned, URIs are no
different from any other way of assigning identifiers to graph nodes.

> > > > >    It *is* all about the graph !
> > > >
> > > > Why should the Semantic Web be restricted to such a limited mechanism?
> > >
> > > Well I do not view it as a limited mechanism.  I have a long standing
> > > challenge, seeAlso (knowledge representation),  that says that I can
> > > represent anything that can be expressed in any language with labeled
> > > directed graphs.  To this day nobody has met that challenge to my
> > > satisfaction.

Well, I think that this is because you allow for arbitrary encodings into
the graphs.  Suppose I claim that strings can be used to representing
anything that can be expressed in any langauge.  How would *you*
distinguish between strings and labeled directed graphs?

> > Well, how do you *represent* - and here I mean represent, not encode - the
> > following first-order sentence using *only* labeled directed graphs?
> 
> I'm just about happy with 'encode'. FOL-in-RDF is like FOL-in-JPEG as far
> as naive RDF (and JPEG) tools are concerned.  They just don't see the big
> picture of the claims being carried through the RDF.

Well then the claim that RDF is a representation language falls down, and
we might as well be using bit vectors to transfer information in the
Semantic Web.

> This 'encode' vs 'represent' is the crux of  the problem, it seems... do
> you have a preferred definition of these concepts?

My view of the difference is that representing something in a formal
language requires having a mapping between the formal language's meaning of
the constructs used in the representation and the meaning of what is being
represented.  So, for example, representing propositional logic in RDF
requires that the meaning of whatever is used to represent disjunction has
a relationship to disjunction itself.  Encodings don't have this extra
condition.

> Dan

peter

Received on Sunday, 25 August 2002 17:34:23 UTC