RE: Provenance in RDF


You might consider how this problem relates to a familiar problem to the museum community.  There, for example, an physical artifact is "discovered" and then over time it may be classified, re-classified, controversially classified, etc.  In effect, there is a constant artifact with different sets of properties associated it by different parties over time intervals (in fact, some of the property assertions have fixed time context (e.g., "Mohammad Ali" had the name property "Cassius Clay" before 1966).

A number of us, Jane Hunter, Martin Doerr, etc., have been working on how to cleanly model the mixed notions of objects changing over time and attribution/characterization of objects over time - Martin coming at this with his more museum oriented perspective and Jane and I coming at this from a more digital library/resource perspective.  

Martin's very complete work is described at  Jane and I have written thoughts on this up in the context of our abc modeling work - see our paper in JODI at

The group of is in involved in a DELOS workshop series try to come up with canonicalized thinking about all this.


Department of Computer Science
Cornell University
Ithaca, NY 14853 USA
Voice: +1-607-255-6046
FAX: +1-607-255-4428

> -----Original Message-----
> From: Dave Reynolds []
> Sent: Wednesday, February 27, 2002 7:14 AM
> To: RDF Interest (E-mail)
> Subject: Provenance in RDF
> We are working on a semantic web related application that 
> needs some provenance
> support. We have various routes for doing this but would be 
> interested in
> hearing of other's experiences. Are there any groups out 
> there that have
> developed applications supporting provenance within RDF that 
> would be willing to
> share their experiences on what worked well or badly?
> To explain a little.
> We are developing a semantic web application for shared 
> information management.
> In this application users are able to attach personal 
> metadata to items and are
> able to view the "soup" of metadata created by many users. 
> For example the same
> item might have many different dc:title fields created by 
> different users and
> the UI should be able to view this data and give response 
> like 'most users call
> this "foo" but one user prefers to call it "bar"'. To support 
> these we want fine
> grain tracking of where the multiple metadata values came 
> from, down to the
> level of individual RDF assertions. The tracking data could 
> include items like
> creator, date and digital-signature, these terms would be 
> defined in a separate
> provenance schema/ontology.
> We are exploring three approaches to doing this - application 
> level, reification
> and out-of-band. Each of these has pros and cons.
> ** Application level
> Treat provenance as a data modeling problem at the 
> application level and
> introduce bNodes to which the provenance can be attached. 
> Thus instead of:
>    subj --pred--> obj
> for any provenanced (is that a word? :-) values use:
>    subj --pred--> <> --rdf:value--> obj
>                      --pv:creator--> "Dave"
>                      --pv:date--> "27/2/02"
> This has the advantage of flexibility and means we can query 
> provenance data
> conveniently using existing RDF query languages (RDQL in our 
> case). However, as
> far as we know this is not a standard idiom and that might 
> make it harder to
> interoperate with other RDF metadata sources.
> ** Reification
> Clearly the official RDF mechanism for representing 
> provenance is to use
> reification and attach the same "pv:*" assertions to a node 
> denoting the reified
> statement.
> This has the advantage of being the standard idiom at 
> present, however the
> uncertain status of reification with the RDFCore WG leaves us 
> nervous. We can
> still query provenance data, though the query would now look 
> rather more ugly
> and verbose than if we take the application level approach. 
> The shear number of
> triples needed is high but (a) is too early to optimize for 
> performance and (b)
> we can in any case hide overhead by implementing a triple 
> store which pretends
> to reify but in fact uses a more compact representation.
> ** Out of band
> In this option we simply make provenance support a property 
> of the API. We don't
> change the RDF assertions in the main fact base at all. 
> Instead we provide API
> calls to attach and retrieve annotations from any RDF 
> assertion. This is related
> to the "quad" notion discussed on this list some time ago and 
> the N3 approach
> that evey statement has an internal context attribute. This 
> has the advantage
> that it hides the mechanics of provenance allowing us to keep 
> the application
> code stable even if the implementation idiom changes. It has 
> the disadvantage
> that we'd need to extend our query support to access this 
> additional API layer
> and is at best unhelpful for integrating with other RDF data sources.
> For our current purposes we will simply pick one and work 
> with it but if anyone
> else has already trodden this path and has experiences to 
> share then we'd love
> to hear from them.
> Dave

Received on Wednesday, 27 February 2002 07:58:01 UTC