Re: Performance issues with OWL Reasoners

>>>>> "KV" == Kashyap, Vipul <VKASHYAP1@PARTNERS.ORG> writes:

  >> There are two different things in the technologies you mentioned;
  >> relational to X mapping tools, and metaschema approaches. They
  >> are quite different. For the instance store, the relational
  >> database is really an implementation detail. It's basically a
  >> reasoner with somewhat limtied expressivity which is persistent
  >> and (hopefully) scalable.

  KV> [VK] Is this true only for TBox reasoning, or ABox and TBox
  KV> reasoning? I had the impression (possibly mistaken) that ABox
  KV> reasoning takes advantage of the relational backend. At least
  KV> this is implied by the following snipper >From an earlier
  KV> e-mail:

The instancestore (or at least the very early version that I
implemented) only does T-Box reasoning technically. The "A-Box" is
stored in a relational database. 

It works like this -- when an instance is asserted (described) the
reasoner is used to localise it's description in the ontology. This
data is then denormalised and put into the database. So, for example,
if you have an entirely asserted hierarchy, you should be able to get
all instances of a given class without (very much) reasoning. In
other, more complex, cases individuals have to be reasoned over quite
a lot at both insert and query phases. 

The reason that it works is because you can't assert relationships
between individuals, so you never need to re-reason things about
instances. For example, imagine these assertions

phil hasSibling martin


This makes me a member of the class of things which have
siblings. Next we assert

martin hasSex Male

Now, my own definition may have changed -- I am a member of the class
of things which have brothers. So the new insert potentially requires
updating our understanding of all instances. But with the instance
store you can't make the first assertion, only things of the form of
the second, so you are safe. 

As I said, this was true of the first instancestore. The current
version is cleverer and can make some assertions of the first
form. You'd have to ask others for details of this. 


  KV> [VK] I guess I need to get out of my laziness and read the
  KV> paper, but how different is the metaschema approach from the X
  KV> to relational mapping approach?

It's a metaschema -- every ontology uses the same schema. For a
relational mapping, you'd expect different ontologies to use different
ones. 

  KV> Even if it is a very different approach, is it able to leverage
  KV> the "scalability features" of an RDBMS enumerated above? I would
  KV> be interested in your responses.


The instancestore uses indexes, yes. This is actually easier -- you
have only one schema, so the appropriate indexes are the same for
every ontology. 

Phil

Received on Tuesday, 19 September 2006 14:16:49 UTC