1 February 2011
Currently active AWWSW members: Michael Hausenblas, David Booth, Nathan Rixham, Jonathan Rees. This report was prepared by JR with help from the others; any shortcomings are JR's fault.
The so-called AWWSW 'task force' was formed at a joint TAG/HCLS meeting 2007-11-05. ('AWWSW' facetiously stands for 'architecture of the world wide semantic web.') The SWHCLSIG (Semantic Web Health Care and Life Sciences Interest Group) was investigating the use of RDF and had encountered a few glitches. Some SWHCLSIG participants came to the TAG with these concerns:
The outcome of the discussion was that an informal group was chartered to discuss "HTTP semantics".
The following is not the only use case, but it surfaces the most important issue. A similar scenario could be constructed around any kind of confusion over what the URI refers to.
Bob composes a document, which we'll call R, and arranges for HTTP responses to GET requests with target URI 'http://example/e' to yield 200 responses carrying R. Among other things, R (and these responses) contain the following:
<div about="http://example/e"> This document, by Edward Example, is licensed under a <a rel="license" href="http://creativecommons.org/licenses/by/3.0/"> Creative Commons Attribution License</a>. </div>
The bit of RDFa means that a document, referred to in R as http://example/e, is licensed as specified. In N3 or Turtle, this would be written
<http://example/e> xhtml:license <http://creativecommons.org/licenses/by/3.0/> .
Alice, reading this RDFa and wishing to republish R, understands that she has a license to republish R. Later, Bob discovers that Alice has republished R and confronts her. (For simplicity in the following, let's assume that Bob and Alice have found a mutually understood way to refer to R.)
Bob: Hey, Edward didn't give you a license to republish R.
Alice: He said that the license applied to R!
Bob: No he didn't, he said it applied to S, the one all this information in R is about.
Alice: But 'http://example/e' refers R, not S! I didn't even know about S.
Bob: No, 'http://example/e' refers S.
Alice: That's ridiculous - you don't use 'http://example/e' to refer to R? That's the customary practice.
Bob: No, nobody told Edward he had to, and anyhow why would you want to refer to R? The interesting document is S.
Alice: How on earth was I supposed to know you meant S instead of R?
Bob: You were supposed to read R. It's obvious if you do that - it wouldn't make any sense if you took 'http://example/e' to refer to R. See, it says right here that <http://example/e> is about penguins, and R is obviously not about penguins.
Alice: But I'm running a stupid search engine. It only looks at the license statement. You can't expect it to read and understand the document so that it knows the URI is supposed to mean S instead of R!
Bob: Yes, I can, it's my URI so I get to use it however I like.
Alice: If I let you play games like this with this URI, how can I ever hold anyone accountable for anything they say in RDF?
Bob: That's your problem. Maybe you're using the wrong technology.
What do you think?
It's easy to poke holes in this story, but the point is the form of the story, not the details. Meaning is burden of proof. If there's no accountability (or as Alan Ruttenberg says, no way to be wrong) there is no semantics.
(Extra credit: If 'http://example/e' refers to S, how would one refer to R?)
In general the 'task force' has been doing three kinds of analysis: empirical (current and current 'best' practice), specification-based, and speculative (future and future 'best' practice).
In all cases the goal is a story that makes sense, an expression of it in some logic (RDF or OWL), and eventually a published set of axioms (i.e. "ontology"). But it is very hard to say anything that makes any sense in this area.
We have postponed consideration of nose-following. The idea is to describe particular methods for using the Web as a 'dictionary' for readers and speakers of RDF (and similar languages) and perhaps nominating one or more as potential 'best practices'. Such documentation would be an operational follow-on to The Self-Describing Web finding and the unexplained figure at the end, and would serve the semantic web and linked data communities and anyone else who cares about automating nose-following.
Documenting nose-following is raised by the new RDF WG charter. As this is a webarch issue and not specific to RDF, it would be appropriate for the TAG or a task force to address it.
The new Web Linking standard is relevant to nose-following, especially given the issues raised by AWWSW. If a URI owner wants to say something about a document, and the page will be untrusted or can't be edited to add embedded metadata, there ought to be some other way to say it, and .well-known and Link: provide this. [refer to previous TAG work here]
Fragids are certainly a concern to those in the group, and as the TAG knows are semantically troublesome, but we have not really investigated them yet.
We have not been addressing this issue directly. Jonathan has been pursuing it separately under TAG ISSUE-50. There is some overlap.
Some in the group have expressed interest in documenting how the semantic web works, but others resist this suggestion. We have agreed to postpone consideration until after we've dealt with the metadata (200) issue.
We have been stuck on the question of how dereference bears on reference. There are two reasons for this. One is the number of 'moving parts' in the combination HTTP + webarch, and the other is the complex set of constraints that seem to be imposed on a solution.
Note that metadata is not always being written by the URI owner. The metadata author may be gambling on stability (no conneg or sessions) and consistency (continuity of content over time). This is their business. However, they will be held responsible for the correctness of the metadata, so the risk of getting it wrong needs to be in inverse proportion to the size of the wager.
Some of our test cases for Web URIs used referentially: (by 'Web URI' I mean those that can be dereferenced, in 3986 terminology - usually GET http: yielding 200)
The moving parts - that is, the entities that could cause someone using a URI referentially to be wrong - include:
The constraints include:
It is quite possible that the problem is overconstrained and some putative requirements must be ejected. It is almost certain that most of the complexity must be ignored or ruled out in any mainstream application involving metadata. But in order to ignore it properly, we have to acknowledge that it is there and understand it. Once it's understood it will be possible to find a way for URI owners and users to express what they committing to or assuming.
A further complication is the "representation / presentation" distinction that Henry documents. So far we have pretended that this doesn't exist. At some point, however, it will be important to make our account resilient to transformations from static to dynamic content.
One might argue that Web URIs are an exceptionally poor notation for referring to any entity whose properties are somehow inferred from what is retrieved. Perhaps there is some better notation available, or maybe one could be designed. The scholarly community uses metadata-based references, which could be rendered in RDF using blank node notation, and also has things like the handle system that support "identifier"-like references that meet their needs. Perhaps the duri: or urn: URI schemes help make certain aspects of the referent clear.
Embedded metadata (metadata that's about what it's embedded in) seems to have a clear subject. In RDF this is usually designated by the URI from which the document is retrieved (i.e. the base URI). This practice is not reliable (conneg, sessions, revision, etc.), but it probably works because the URI is interpreted as intended. [To think about: thismessage: .]
However, embedded metadata is a special case, and will not usually help us indicate the subject of metadata in general. The metadata itself might do so of course, but the common practice of using a URI also does.
There are both sunk cost and aesthetic arguments against rocking the boat and in favor of plugging its leaks. The investment is not just in deployed metadata and its processors, but in the development of frameworks (such as webarch) and specifications that have encouraged the status quo.
Another reason to forget about it might be if a competing (i.e. non-interoperating) use of these URIs took hold, as has been discussed recently on www-tag (see thread starting here; also this from Harry Halpin). But this makes little difference as some other way would have to be found to refer to these entities, and then that would have to be explained.
The idea that Web URIs are unsuitable for reference and ought to be unstudied is belied to some extent by the fact that the community continues to use them referentially quite happily. The contrast between the lack of specified semantics for Web metadata and its usefulness and importance is a puzzle that needs explaining.
Fortunately, anything we learn about Web URIs can be applied to any notation used for this purpose, since all such notations run the risk of most of the same failure modes as Web URIs. Exactly what handle and DOI 'owners' or users can be held to is not spelled out very well; Crossref's member rules are quite rigorous, but leave vague what the constraints are on what web servers are supposed to do (the landing page issue described above applies, as does conneg confusion). duri: URIs and URNs, even when they're actionable, are just as confusing in the presence of content negotiation, session sensitivity, and so on as are Web URIs. By understanding better how Web URIs are used referentially, and the circumstances in which Web metadata is and isn't successful (not just technically but socially), we ought to be able to design new systems and improve the ones we have.
We usually phrase the problem in terms of HTTP but the general question is not semantics of HTTP but semantics of any URI that is used in a similar way - that is, the URI in the REST/webarch/3986 abstraction, not the one protocol specifically. The AWWSW problem has to explain web architecture first, and its instantiation in HTTP second.
Metadata vocabularies that we've been keeping in mind include Dublin Core, FOAF, the Information Artifact Ontology (IAO), RDFS, and CC REL. Genont is interesting but we're not aware of any deployment. FRBR provides a useful analytical framework and we use it as a reference point from time to time. BIBO is related but does not seem to use Web URIs referentially.
Tactically, we have decided not to worry about time for now, as it is an unsolved problem in the RDF community in general, and we have nothing in particular to contribute. There is enough to worry about with instantaneous confusions.
Members of the group come in with widely varying approaches to semantics and the nature and purpose of RDF, and this has been a source of friction.
Work to date includes
The TAG has been tracking AWWSW's work under ISSUE-57 (redirections). It might continue this out of inertia, but most of what's covered would make more sense under ISSUE-63 (metadata architecture).
Thanks to Alan Ruttenberg, Stuart Williams, Pat Hayes, Harry Halpin and others for comments as this report was being prepared.