Re: [xml-dev] Meta-somethingorother (was the semantic web mega-permathread thing)

[Jonathan Borden <jonathan@openhealth.org>]

Michael Rys wrote:

Once somebody shows me a description logic inference engine that actually scales, I will become more interested in this aspect.

DL inferencing is relatively slow although boundedly so, however, for the types of problems it addresses, it may be a necessary cost at this stage of the game -- note that the DL dialect of OWL is selected because it is thought to be *faster* than the completely unconstrained RDF dialect of OWL: OWL Full.

The one advantage RDF as any extended entity relationship model has over tree models like XML, is the ability to represent relationship graphs. However, for that, there is a very well suited tuple based model: the relational model. Having more semantics associated with the relationships is useful, but only as long as the relationships are built-in.

Best regards

Michael (who in his former live implemented a hybrid database/description logic system)

See that's exactly as I see it, and others have/are currently worked on/are working on/ this approach: a hybrid RDB inferencing approach. This is really what data mining applications *are*.

*The* big initial driver of OWL was DARPA, as it is well known that OWL was incarnated as the "DARPA Agent Markup Language" and it is well known that there is alot of interest in certain well-funded government circles regarding threat classification, pattern recognition. For example suppose we have gobs of random information, and from these gobs we have a way to correlate the information into individual collections, for example, suppose we can correlate a whole host of phone conversations as involving individuals. Now suppose we have lots of other types of information that involve other (as yet unnamed) individuals. What we want are a set of inferencing operations that will allow us to *equate* individuals identified by sets of phone conversations with individuals identified by financial transactions. Get the picture.

In OWL, these there the operations that the inferencing engine can spit out:

:Joe owl:sameIndividualAs _:1

where _:1 identifies a blank node.

formally known as graph merging, informally known as smooshing.

Now I can give similar examples from medical diagnosis.

In any case all of the above examples would typically run off of a relational db platform. DL is then seen as primarily a type of data mining operation.

There is more of course, OWL/RDF is web aware in the sense that all names are URIrefs, and so there is the capability of linking information across platforms. Now *iff* RDBMS had ever actually become platform independent there might not ever been a need to develop a new schema language -- e.g. you might have URIified SQL names and run your PIVOTS across machines in a transparent fashion, but the current state of RDBMS is hardly that. The more you encourage SQL developers to run C# code in the DB, the less you are using the relational model, the the point where tuples may become somewhat irrelevant. So from an entirely *practical* point of view RDBMS are not transparent to the Web. Oh well, that's life and business. It is very similar to the fact that genetic evolution is not always the most efficient but it is what it is. So goes the evolution of technology: If there are problems to be solved for which the current tools don't quite do the job, sometimes we write entirely new tools rather than adapt the ones we have, and then the old tools get rewritten with the new features etc.

See, you would have been a perfect person to serve as MS representative to WOWG if MS had chosen to get involved in that effort. Peter F. Patel-Schneider who did alot of the heavy lifting for OWL certainly has a fairly extensive knowledge of RDBMS.

Jonathan