Re: rdf inclusion

Jeff's response to Pat's "dream world" comment reminded me that I
hadn't responded. 

     [Jeff Heflin]
     >In terms of merging RDF graphs, daml:imports means you can't add that
     >triples from some graph unless you also add all the triples from the
     >graphs of the resources that are imported.

   [Pat Hayes]
   But wait a minute.  What does it even mean for your ontology to say 
   what my reasoning engine can or cannot do? Of course I CAN add 
   triples from one graph without adding triples from another. All that 
   any ontology can do is to express some propositional content. What 
   another engine does with that content can be reasonably expected to 
   conform to the semantics of the language, but that's about all. If 
   the engine decides to ignore some of what you say, that's it's 
   business, not yours. Ignoring part of any RDF graph is perfectly 
   valid considered as an inference, after all: an RDF graph entails all 
   its subgraphs.

I would disagree with Jeff on technical grounds: the daml:imports
statement states what is *not* to be included, as opposed to what
*must* be included.  That is, it points to a DAML+OIL document, and
says that any triple outside this document is not encompassed by the
import unless specifically authorized by daml:imports within that
document.  As I've said elsewhere, this leaves open the possibility
that the document uses proper-name URIs from other, unimported
documents.

However, pragmatically, Jeff has a point: If my inference engine fails
to draw an inference, and the inference follows from premises I have
in hand plus stuff from an imported ontology, I can't call up the
tech-support staff for the ontology and expect much if the
conversation goes like this:

   Me: "My inference engine didn't conclude P when it clearly follows
      from premises Q1 and Q2.  Q1 is in my local dataset, and Q2 is
      in your ontology, which my dataset imports."

   Tech Support: "What exactly does your inference engine do with Q2?"

   Me: "Oh. It ignores it.  Pat Hayes said I was free to do that, and
      it did make the implementation of my inference engine much
      simpler." 

   Tech Support: <dial tone>

   [Pat]
   I think this entire discussion is in a dream world. First, there are 
   no clear notions of definition to appeal to. 

And none needed.

   Second, no ontology can restrain the actions of a remote inference
   engine.  

True but irrelevant.

   Third, why would one want things to be different?

Because it's not a matter of "restraining" anyone.  If ontologies are
posted with well-defined semantics, and people advertise inference
engines that claim to use those ontologies, then consumers should have
a reasonable expectation that the inference engines respect the
semantics and draw mostly useful inferences that mostly don't require
an exponential amount of work.  It's in everyone's interest to play by
the rules.  Am I missing something?

                                             -- Drew McDermott

Received on Wednesday, 8 May 2002 20:29:57 UTC