Re: Social meaning discussion 6th March

>* Tim Berners-Lee <timbl@w3.org> [2003-02-28 15:34-0500]
>>  Dan Brickley wrote:
>>
>>  >* Tim Berners-Lee <timbl@w3.org> [2003-02-28 11:47-0500]
>>  >
>>  >
>>  >>I am concerned that you have thrown out the baby with the bathwater.
>>  >>And still left some bathwater. ;-)
>>  >>Our views do seem rather different
>>  >>
>>  >>What is required, and easy, is to say what an RDF document means.
>>  >>
>>  >>What is not required and a bad idea is to explain how to use it.
>>  >>
>>  >>1. The meaning of an RDF document is that of the statements.
>>  >>2. The meaning of the statement is defined by the definition
>>  >>of the predicate

What is a "definition" of a predicate? Nothing in any of the RDF or 
OWL specs refers to definitions.

>, as applying to the subject and object identified by the
>>  >>definition of the subject and object terms.
>>  >>

BTW, there is a huge hole here: Nothing in any W3C spec, except 
possibly that for URNs,  provides any account of what it means for a 
URIref to "identify" something. There aren't any names on the Web; 
or, more accurately, there are millions of them, but there are no 
protocols or standards for specifying what they are names of.

>  > >>
>>  >
>>  >This for me is the crux: do we mean the machine oriented 'definition'
>>  >in RDFS or OWL or N3, or some more rounded/scruffy/social notion of
>>  >definition.
>>  >
>>  >
>>  Whatever is available.  It tends to depend on the term.
>
>Yes, there are a variety of strategies for representing the meaning of a
>term. The RDFS spec tries to acknowledge this in
>http://www.w3.org/TR/rdf-schema/#ch_comment where we say:
>	"""A variety of documentation forms can be combined to indicate
>	the intended meaning of the classes and properties described
>	in an RDF vocabulary. Since RDF vocabularies are expressed
>	as RDF graphs, vocabularies defined in other namespaces
>	may be used to provide richer documentation."""
>(I suspect we could usefully lose the word 'intended' there;
>  and I certainly wouldn't replace it with 'social'!)

Please do not lose the word 'intended'. If you omit it, that 
paragraph needs to be struck from the document. The whole point of 
distinguishing between 'intended meaning' and (actual) meaning is 
that the latter is often not the same as the former. Comments can 
indicate intended meanings; they can be documentation. They cannot 
constrain or specify actual meanings.

SO, to return to the point, the above is NOT about specifying the 
meaning of a term.

>  > >Use case: I want to add 'foaf:uncle' to my FOAF vocabulary at
>>  >http://xmlns.com/foaf/0.1/ and intend to express as much of the
>>  >conventional
>>  >meaning of 'uncle' as I can with the technology available, ie. RDFS +
>>  >OWL + N3 + HTML/prose. In the case of 'uncle', most of the meaning is
>>  >invisible
>>  >to RDF/S and OWL. But most of it could be handled by N3 rules, assuming
>>  >we had foaf:parent, foaf:sibling, gender vocab etc.
>>  >
>>  >
>>  Well, that is a big assumption.    So let's be sepcific
>>
>>  { ?x :uncle ?y } log:iff  {  ?x   :parent  [  :brother ?y ] }.
>
>(OK, lets forget the 'husband of aunt' case for simplicity.)
>>
>>  is a precise definition of the term for someone for whom parent and brother
>>  are defined.  But this of course doesn't really help us as somewhere the
>>  thing has to be grounded in english.

That really is a persistent and very harmful delusion. The entire 
point of the whole AI/KR effort, of most applied ontology work, and 
even of a large part of commercial relational database technology, is 
that large sections of meanings do NOT need to be 'grounded' in 
English. If all meanings had to be grounded in English then the SWeb 
would be a non-starter. Fortunately, this isn't even remotely true, 
which is why the SWeb isn't totally impossible.

>  >
>>  '"?x :uncle ?y" indicates that y is the uncle of x'
>>
>>  is an english definitoin which will do for a lot of people.
>
>Yes, though I do hope at some point a WG offers some best practice
>guidelines to authors of rdfs:comment or owl2:comment properties to
>help them write their comments in such a way. Imagine if they wrote
>"'xyz(?x,?y)' indicates that there is an xyz relation between ?x and ?y".
>Or variations on that theme. We can I think do an incrementally better
>job of helping people be clear about the meaning of the terms they create.

There is a GREAT danger here of having English comments be wildly 
overoptimistic. In many ways it is better to have the  "indicates an 
xyz relation"  form rather than something which claims some 
inexpressible content or grossly oversimplifies some actual meanings.

This isn't just griping: we have about two decades experience, and a 
pretty solid collection of cases studies, of just how misleading such 
naively optimistic 'English glosses' of intended meaning can be, and 
how they lead to errors, cause failures of interoperability. and even 
dangers. Its often called the 'gensym fallacy'.

>This view requires that we conceive of the possibility of there being a
>gap between term meaning and what definition_s_ of the term say, especially
>as there may be various definitions, and these may be refined and
>improved over time. The convenient idealisation is that there is a
>single meaning towards which all these definitions are fumbling...

That is a dangerous and false assumption. It is based on a hopelessly 
naive view of what 'meaning' means in normal human discourse; but 
even a moderately competent lexicographer could tell you that there 
is almost never a single meaning for almost any term in almost any 
language. Worse, the more exactly you require the meanings to be 
specified, the more fragmentation of meaning you get. For example, 
many dictionaries distinguish two or three meanings of 'bank' 
(financial institution, edge of river, aeroplane turn). BUt if you 
start getting fussy about distinguishing senses, the first of these 
can be used to mean a company, an agent, a legal entity, a building, 
a location, a collection of buildings, even things like an 
architectural style. Most common word meanings fracture into about a 
dozen meanings when 'meaning' is required to be delineated with any 
precision. And our ontologies REQUIRE that meanings be used with this 
kind of precision, because they have taxonomies of 'kinds of entity' 
built into their very skeletons, so they are obliged to be achingly 
precise about distinguishing locations from agencies from financial 
institutions, etc.,.

>
>>  I would expect a good spec to have both.
>
>I very much agree. Belt and braces. HTML, RDFS, OWL, (rules, ..., ...).

If the SPEC refers to meanings defined in English then it needs to 
say what it means by that. And if it says anything that implies that 
any software depends on that meaning, then it is guaranteed to be 
useless, because to write such software is beyond the competence of 
anyone at the present time.

>
>>  The formal  information is a useful axiom.
>
>Yes. And I believe many of us anticipate there being languages in 2-3
>years time that will make such formal information more widely accessible
>within RDFS/OWL/+++ vocabulary descriptions.
>
>>  >So when you say 'the definition of the predicate', that to me is where
>>  >our problem lies. (I'm happy btw having predicates drive the meaning
>>  >of statements, though I think this is something we stipulate rather than
>>  >show). Which 'definition' of the predicate sets the meaning?
>
>>  One would expect that any maths given with an english definition would be
>>  consistent with it.
>
>A reasonable expectation,

But try checking it out in reasonable-sized cases.  (Listening to you 
talk about this stuff is like listening to a bunch of guys in a bar 
discussing how to climb Everest next weekend.) In any case, the issue 
before us is, which is normative? If the meaning of the English is 
declared to be normative, then RDF is dead in the water, since it is 
immediately totally impossible to write any RDF-conformant software.

>and one that indicates something we've not
>really explicitly articulated: that we (well, maybe 'we') believe
>as our guiding fiction that class and property terms have a meaning
>that is independent of the languages (html, rdfs, owl, n3, ...) that
>we use to try to characterise that meaning.

The sooner you stop being guided by that illusory fiction the better. 
Its a very, very, very poor guide. It can't even handle the current 
state of RDF and OWL.

If you think about it for a second, its kind of obvious that it 
doesn't even make sense. Did you know what "Froxible" means? It's a 
word in Fragolian, and maybe you aren't familiar with Fragolian...oh, 
but wait: for you, meaning is *independent* of language, right? So 
*you* ought to know what it means even without knowing the language.

>When all goes well,
>those varying characterisations are in sync. When things go badly 
>(discuss...:)
>
>
>This is somewhat related to RDF's practice of treating properties
>and classes as just more entities from the domain of discourse: they have a
>life beyond the RDFS assertions that describe them, or the OWL-encoded
>rules that describe them slightly better. There are, to stick with
>my use case above, facts of the matter about the proper use of
>the 'foaf:uncle' which we attempt to capture using whatever tools are
>within our grasp. Meanwhile, the property gets deployed and used. This
>division of labour allows people to get out there and _use_ RDF
>while W3C WGs work on better ways (OWL being the latest but I doubt
>the final) of characterising the meaning of these terms.

The meaning IS the use.  The best that the W3C WGs can do is to 
provide the world with more useful languages and hope that it starts 
using them. But the W3C is not a fount of authority about meanings. 
To set oneself up as the meaning-shop would be to invite 
well-deserved ridicule.

Pat


-- 
---------------------------------------------------------------------
IHMC					(850)434 8903 or (650)494 3973   home
40 South Alcaniz St.			(850)202 4416   office
Pensacola              			(850)202 4440   fax
FL 32501           				(850)291 0667    cell
phayes@ai.uwf.edu	          http://www.coginst.uwf.edu/~phayes
s.pam@ai.uwf.edu   for spam

Received on Sunday, 2 March 2003 01:29:13 UTC