Re: What is assumed?

>Pat,
>
>Some thoughts on your comments about flower traders are below.  I 
>also note that your latest semantics draft addresses some of points 
>I mention.
>
>At 02:32 PM 7/26/01 -0700, pat hayes wrote:
>>>At 12:19 PM 7/26/01 -0700, pat hayes wrote:
>>>>I think this example has been valuable because it illustrates a 
>>>>much bigger and more important point than the one we started 
>>>>with. The implied meaning to an English speaker is *completely 
>>>>irrelevant* here. RDF is supposed to be used by software agents, 
>>>>not English speakers (if we could rely on the native savvy of 
>>>>English speakers, we could do it all in HTML). So the only 
>>>>content that any piece of RDF (or DAML, or KIF, or any of these 
>>>>'formal' languages for capturing content) actually contains is 
>>>>what some mechanical agent could infer from it (together with 
>>>>whatever other pieces of RDF it is able to glean from various 
>>>>sources, of course.)
>>>>
>>>>If you apply that criterion to Brian's example, and if you take 
>>>>it to be an assertion, then all that you could possibly infer is 
>>>>that four things exist and a few relations are true between them. 
>>>>If this is supposed to convey the fact that one of these things 
>>>>is a 'service' and therefore that it implies that a lot of other 
>>>>things exist (eg batches of roses which are available for sale or 
>>>>purchase -I've lost track of whether we are selling them or 
>>>>buying them), then something needs to say that.  Nothing in 
>>>>Brian's example seems to say that, however, so there is no basis 
>>>>for anything to be able to conclude it. (I would hazard a guess 
>>>>that the only way to say what needs to be said here is to use a 
>>>>universal quantifier, by the way.)
>>>
>>>Yes... I think that the examples were incomplete in that a 
>>>definition of specific symbols used (to be guessed at from the 
>>>names used) was needed.  Without this, there was no logical 
>>>distinction between the buyer and seller examples (modulo a small 
>>>change of quantity values).
>>>
>>>But, at some point, a computer program has to interface with the 
>>>real world.  Maybe the only difference between the buyer and 
>>>seller processes is that the results of invocation of the service 
>>>are delivered to different printers.  It may happen that one 
>>>printer is in a warehouse with a staff that reads the requested 
>>>quantities and other details, picks the desired roses off the 
>>>shelf and puts them in a van for delivery to the stated address. 
>>>That would be the seller service.  Another printer might be 
>>>serviced by a staff who pick up the details and send their van to 
>>>a designated address to collect some roses.  That might be the 
>>>buyer service.
>>>
>>>How much of this external-world difference do we need to encode in 
>>>the computer program?  It may be that the only difference between 
>>>the buyer and seller, as far as the program is concerned, is an 
>>>associated port address that selects an appropriate printer to 
>>>receive the service instruction.  So if the previous examples were 
>>>augmented with something like:
>>>
>>> buyer usesPrinterPort "1" .
>>>
>>> seller usesPrinterPort "2" .
>>>
>>>that states a logical difference between the services.
>>
>>Well, no. Of course programs rely on such things all the time, but 
>>my point is that in cases like this the information is NOT being 
>>conveyed as a 'logical difference', ie in this context, in the RDF, 
>>but in some other way that is invisible to an RDF processor. Which 
>>is fine, as long as we face up to that. In the meantime, however, I 
>>am assuming that what we are supposed to be focussing on here is 
>>what can be said in RDF.
>
>OK, it's not a *logical* difference.

I shouldnt have used the L-word. The issue is, is it expressed in the 
RDF somewhere, or not? If not, then my next question would be, how 
then is it being conveyed between the people who are 
writing/publishing/reading this RDF? If the RDF can only be 
understood by virtue of some other content represented somewhere else 
in some other format, then it seems to me that (1) it - the RDF - 
isn't doing its intended job of representing Web content in 
machine-processable form, and (2) there really isn't any point in 
anyone using RDF at all, since we could convey the content using HTML 
(as everyone does now) on the same basis, ie so that it can be 
understood with the help of 'external' conventions about meanings - 
in the case of HTML, those might include the conventions of all the 
world's natural human languages.

>
>>>But the context within which the services are deployed (defining 
>>>an interpretation?) is part of the situation in which we care 
>>>about the results of the computation.
>>
>>If we were programming local computations in a limited, known 
>>context, then we could rely on all kinds of assumptions. But RDF is 
>>supposed to be publishable on the Web, right? What 'context' can be 
>>assumed between a publisher of a web page and a reader of that web 
>>page, beyond the protocols used to transfer the information?
>>
>>>Maybe this goes against normal practice for model theory.  I 
>>>notice that you said in your strawdog note that common model 
>>>theory practice is to make the available domain of interpretation 
>>>as general as possible.  I suppose that the vital thing we need is 
>>>for programs to exchange information that is  constructed using 
>>>the same assumptions about the interpretation.
>>
>>If there are any global assumptions then by all means let us build 
>>them in. In the strawdog, for example, I assume that literals have 
>>a fixed global interpretation which never varies, so anyone using a 
>>literal can assume that any reader will have the same 
>>interpretation. We can have fixed interpretations for things like 
>>rdf:subject (in fact for anything starting rdf: or rdfs:). But how 
>>do we establish globally fixed meanings for things like 
>>wholesaleSupplierOfRoses ?
>
>I think that a possible difference of RDF/DAML/etc from traditional 
>maths/logic is that the names used (URIs) do have a globally fixed 
>meaning, even if we don't know what they are...

Ah, but you can say the same about math logic. It's that "knowing 
what they are"  that is supposed to be provided by publishing the 
content in RDF (or KIF, or DAML, or whatever.)

>I came into this whole area from the direction of content 
>negotiation between software systems (in particular, between email 
>user agents).  I have all along held a view that the mechanisms of 
>content negotiation (i.e. form of description of content handling 
>features, rules for selecting data that can usefully be transferred, 
>etc.) and at this level I fully concur with your view of finding 
>universal mechanisms that don't make any assumptions (e.g. about the 
>features being negotiated).  But to actually use this in practice, 
>particular features have to be assigned names for use in the 
>negotiation process, and a user agent must be imbued with knowledge 
>of what capabilities it has: how they are identified and what they 
>mean in terms or utilizing the software and/or hardware capabilities 
>of the agent.  Such instance knowledge about an agent cannot be 
>captured in the theory of negotiation.
>
>So, I think there are two parts to a system for content negotiation: 
>the common rules of negotiation which are based on explicitly 
>defined knowledge, and the implicit knowledge (human-defined, 
>hand-coded, etc.) that systems have of the symbols for things that 
>they deal with (which you describe above as being "invisible to an 
>RDF processor").  I see a  goal of our work being to maximize the 
>explicit element, so it can be re-used across a widest possible 
>range of applications.  This corresponds to things described by a 
>model theory.  But I find I cannot conceive of a practical system 
>that can completely dispose of implicit knowledge (which I think 
>corresponds to a particular interpretation).

I agree (Well, particular SET of interpetations, in general). There 
is implicit knowledge even in the use of RDF itself, eg the fact that 
the middle thingie is supposed to be a property. But part of the 
utility of having a model theory is that it focusses on the content 
that is explicitly encoded in the RDF syntax; the part that is 
explicitly stated by the RDF itself; and all the rest. The rest has 
to be conveyed some other way, maybe by a kind of human coder's 
conspiracy to wriote code on the basis of a shared assumption.

But now, go back to the example, where someone is publishing some RDF 
to the world in general in order to attract flower-buying customers. 
Where is the extra meaning (about the flower-trade specific URIs used 
in the RDF) encoded? Since this is being published to the world in 
general, it had better not rely on any flower-trade-specific 
assumptions, or else someone in some other trade might misinterpret 
it in some potentially dangerous way. We can hardly assume that the 
URI itself encodes anything about flowers in particular. That seems 
to leave the RDF.

>Returning to the 'wholesaleSupplierOfRoses', and Brian's example, 
>the knowledge of buying and selling of roses is implicit.  Further, 
>I don't think it is possible to come up with any example of a 
>practical, useful system that does not depend on some level of 
>implicit knowledge.

Well, thats a reasonable position, but how do you reconcile it with 
the idea of world-wide broadcasting of content?

>But with RDF, used as intended, the implicit knowledge associated 
>with any URI symbol does not depend on the particular context, 
>because URIs are defined to be globally unique.  In model theoretic 
>terms, I suppose this means that there is only one interpretation 
>for an expression, even if a processor may not know what it is.

No. The global uniqueness is of the name, not of the interpretation 
of the name. I don't think that URIs as a class have *any* general 
implicit knowledge attached to them, in fact. (URL's do, but that's 
not a kosher term any more, I gather.)

>
>>>The fewer such assumptions, the less room there is for a 
>>>communication to be misinterpreted.  But I'm pushed to see how we 
>>>can make a useful program without many assumptions about the 
>>>environment it communicates with.
>>
>>I thought that was the central point of RDF. It certainly is for DAML+OIL.
>
>I do not view the boundary between explicit knowledge and implicit as fixed:

Not in any absolute sense, but each formal langauge chooses a place 
to draw its line, and those lines are all quite clear (though not all 
in the same place.)

>in a sense, much of maths, and computing practice, has been about 
>finding ways to codify and generalize specific knowledge and apply 
>the results more widely.  I view RDF/DAML/etc as pushing that 
>boundary, but I don't see the goal of completely eliminating 
>implicit knowledge (assumptions) as being achievable.

Well, RDF/DAML hardly push the boundary. In expressive power they are 
way behind conventional logics or Krep schemes that have been used 
for years in AI and database technology. But I digress....

Pat

---------------------------------------------------------------------
(650)859 6569 w
(650)494 3973 h (until September)
phayes@ai.uwf.edu 
http://www.coginst.uwf.edu/~phayes

Received on Monday, 6 August 2001 20:48:44 UTC