- From: Tim Berners-Lee <timbl@w3.org>
- Date: Mon, 27 Jan 2003 18:16:50 -0500
- To: Bill de hÓra <dehora@eircom.net>
- Cc: Jonathan Borden <jonathan@openhealth.org>, David Booth <dbooth@w3.org>, www-tag@w3.org
On Saturday, Jan 25, 2003, at 08:29 US/Eastern, Bill de hÓra wrote: > > Let me be clear about my intent. I would much prefer a many to 1 > relationship. The important bit being that there is 1 referent (if > there was only one one URI in the relationship, I'd argue that we > wouldn't a seamntic web at all). Indeed I think at certain levels you > /must/ have a many to 1 relationship. Over the years I've seen people > get mighty annoyed that it could be any other way, including myself. > But what I haven't seen is an iota of evidence, or any indication that > this relationship could ever hold on a global system where machines, > not people, are doing the heavy lifting of intepretation. It could I > believe, be held in a closed world, with not insignificant > coordination costs. IMHO, in denying denotation by use, we might as > well be denying the presence of latency, or partial failure. My point > is that many to one is not an observable truth, but as likely as not > the result of a processing or filtering step. > You observe denotation by use in natural language systems. The HTTP space is based on DNS, where denotation by use is *not* observed. The worldwide coordination system one requires to make this happen has been established already, and the semantic web is taking advantage of it. It includes a delegated authority name space, a social system to manage its operation, and a network protocol to resolve in around a second the authoritative answer to the question, whether it be at the DNS level or at the level of an RDF thing defined in a RDF document in HTTP space. So, take heart! It works. Its importance (and the dangers of disrespect to this system) are of course great, and the TAG has spent some time figuring out how to explain to people that URIs (with fragids) do identify just one thing and tehy > Now, the TAG can say that 1 to 1 is canonical or axiomatic, but so > what? We can't even enforce this with authority based URIs - different > authorities can own URIs at different times. At the moment, that is at least a theoretical problem, and one of the tasks the community has is to ensure a greater and greater trust that domain names will be stable. I have written quite a lot about this. In practice, though, to use the semantic web, there has to be a trust that the domains will be stable though the life of the project. (Looking over a periods of 100s of years, I would hope we still have stability, but also librarians will have history-aware systems which operate with an understanding of the time-variation of domain names, and take that into account when interpreting documents. But that is a rat hole discussion right now) > An axiom does not eliminate the potential for multiple referents. > Defensive coding against garbage input, because the intent of the > system is that there can be only one. > > We can build a semantics to state, clearly, that a URI has only ever > one referent. But so what? The coordination costs in doing so are > surely untenable for a global system (which is why I keep using > backlinking as an analogy). Bad analogy. You should use forward linking as your analogy. Forward linking works. No one needs to find all occurrences of an identifier for the system to work. There is no requirement for global consistency of the semantic web. Consistency is required among the working data set of each agent. > By the way we already have such a semantics in RDF! RDF acknowledges > denonations must be distributed in an interpretative step before > inference can begin. However, if the 1 to 1 relationship had anything > to do with the web, that part of RDF's Model Theory could be dispensed > with as redundant. > > A semantics does not eliminate the potential for multiple referents. > Defensive coding against garbage input is needed, because the intent > of the system is that there can be only one. Defensive coding is certainly a good idea. When it finds an inconsistency, the response should be "One of these is wrong" rather than "Oh, this URI identifies two distinct things". > On the other hand, if the TAG said something along the lines of Tim > Bray's response to my request, I could start to apply techniques to > disambuguate referents to code's mainline. The top layer of the system > would still be driven by logical reasoners, things like OWL and RDF > are in no risk of getting canned if that's what people are worried > about. But the dismbiguation layers between inference machines (RDF > reasoners) and the data machines (web servers), would be open to some > badly needed experimentation. You know, keep the 1 to 1 model, if it's > perceived to be that valuable. But at least acknowledge there is a > layer missing in the cake between web servers and inference machinery > that does dismabiguation. That layer is neccessary if only because > things like RDF and OWL break when multiple referents are present in > the data. Today maybe we don't see it, because we are that layer. > > Defensiv coding is removed purely by relaxing the tolerances being > engineered to, because the intent of the system is that there might be > more than one. > > Cool URIs do change. That's why you had to remind us they don't. > (BTW, I meant "Don't look for a domain name that sounds cute. Just make sure you don't keep changing it. That's what it really cool". It was in days when everyone was sporting these vanity URIs like http://sock.it/to/me and whatever.) > Bill de hÓra > >
Received on Tuesday, 28 January 2003 19:25:16 UTC