US-VO Semantics: from Bertram Ludasher
ludaesch at SDSC.EDU
Tue Sep 24 08:09:46 PDT 2002
>>>>> "CG" == Carole Goble <carole at cs.man.ac.uk> writes:
CG> Bertram wrote in his discussion:
>> If you need real *reasoning* capabilities, i.e., given a definition of
>> a concept A and another definition of a concept B, does one contain or
>> exclude the other (or are they equivalent), then you need to look at
>> "Description Logics". Probably you do not need this functionality, but
>> I just mention it for completeness.
CG> I find this extremely strange as a comment. Bertram suggests that
CG> there are the "before semantic web" and "after the semantic web"
CG> languages. This is not true. OWL (see point (b)) is heavily influenced
CG> by Frame-logic work by Karlesruhe. RuleML (see (c) ) is linking Datalog
CG> and Horn Logic with DAML+OIL. To be so dismissive doesn't do you
CG> justice Bertram!!
Oops.. didn't want to imply that there are before and after the
semantic web languages! (in fact, I tried to locate in which context I
said the above -- but sure: these are my words ;-)
What I tried to say above is that for the application at hand, it may
or may not be necessary to have *real* reasoning capabilities a la
Description Logics. Often time (at least in applications we have
encountered), a "simpler" reasoning a la Datalog is good enough:
Essentially formula evaluation over a given database
instance. However, sometimes you need the reasoning machinery because
you want to know whether some concept definition implies another, or
is consistent with another etc. Another quite hot application of
reasoning is of course in semantic query optimization.
Another comment I made at some point and that may (but hopefully
doesn't) upset the community is that there are several formalisms out
there that co-exist but it is not clear which one will have the
greatest impact. For example RDF, RDFS vs DAML+OIL. In some sense
they are complementary, but in another they are competing.
Don't get me wrong: I'm a friend of the Semantic Web! However, in our
data mediation projects, we are still wrestling with some query
processing issues, that's why we have still a few months to go before
we look very closely at the state of the Semantic Web then. (In the
XML querying world we made the same experience that things are in a
flux for a while)
CG> (a) The classification based reasoning that DLs give you is EXACTLY the
CG> kind required to build controlled vocabularies. Controlled vocabularies
CG> for content will be the primary use of ontologies in this field I
CG> suspect as they are in biology. we are using DAML+OIL to rebuild the
CG> Gene Ontology, and I cantell you we aren't doing it for our health. We
CG> are doing it because "we DO need this kind of functionality" to build
CG> large, multi-axial, coordinated and multi-authored ontologies.
CG> Subsumption based reasoning, and its associated query inclusion and
CG> indexing is highly valuable, not to mention the axioms for ensuring
CG> concept satisfaction. I'll put a couple of papers about our experiences
CG> on this on a web site -- they are in press so I have to seek permission
CG> to do so. We have also used these approach in myGrid for describing and
CG> matching services (http://mygrid.man.ac.uk/rpapers.shtml), as have HP
CG> (see http://www2002.org/CDROM/refereed/211/index.html)
CG> All of this is reasoning with *intension*.
CG> If you are arguing about inference over instances with variable, then
CG> that is another argument. *extensional* reasoning is better served by
CG> conventional database languages and even deductive databases,
yes. and again: my comment wasn't meant to dismiss inference at all
(in fact, I'm an old time inference guy myself having worked with
deduction systems and tableaux reasoners of various sorts).
For example in our ICDE'01 paper I explicitly suggested the use of
description logic to register new concepts relative to an existing
ontology. However, not always does one need reasoning. But I agree,
multi-authored ontologies is a clear case *for* reasoning!
CG> For controlled vocabularies Deductive Databases seem like an overkill --
CG> this goes back to your point about "what do you want to do with your
CG> representation". In fact, DD for ontological reasoning is probably
CG> missing the point.
I agree that DD would be not appropriate for reasoning with controlled
vocabularies. But not because they are too powerful, but because they
are too weak.
CG> b) DAML+OIL and OWL (the language that has derived from DAML+OIL,
CG> proposed by W3C WebOnt and now released in draft form) does have the
CG> benefit of a community of tool builders producing parsers, editors,
CG> matching engines, viewers, annotators etc etc. This should not be
CG> underestimated. It also gives you a range of reasoning engines, if you
CG> want that sort of thing, or just a common exchange language if you want
CG> that sort of thing too.
CG> I recommend that you take a look at the OWL spec:
CG> By the way, OWL-Lite is (I believe) supportable by F-Logic reasoners as
CG> well as conventional DL reasoners.
very good -- I haven't stopped by there recently...
CG> c) The use of Datalog style reasoning with deductive rules is useful, of
CG> course, in particular if one wants to do process modelling and planning
CG> & scheduling work. Attempts are being made in the Semantic Web world to
CG> incorporate DAML+OIL terms in rule languages. An interesting place to
CG> look at is RuleML activity:
CG> http://www.dfki.uni-kl.de/ruleml/ I haven't used FLORID in a
CG> long time but I do remember problems with robustness and legacy data,
CG> and wasn't it undecidable as well as intractable? Tractability is only
CG> obtainable in DD by the Closed World Assumption. Again, CWA can be
CG> presumed over querying instances, but reasoning about possibilities, as
CG> one would want to to do ontological intensional reasoning, is best
CG> served by the open world assumption.
Indeed F-logic (the FLORID and the FLORA implementations e.g.) like
Prolog is Turing complete. However there are decidable fragements. But
then again, one (well, at least we) don't use it for reasoning but for
"formula evaluation" over a given instance. Of course using function
symbols you can reduce reasoning to a query evaluation problem (see
the paper by David Toman in last year's DL workshop). For specialized
reasoning tasks, in particular with description logics, there is no
doubt that specialized reasoning engines are the way to go and that
they clearly outperform other approaches. A perfect example is FACT!
CG> d) RDF, though a simple graph model, is weak but it is promising for
CG> data integration and does have some reasoning capabilities. The point of
CG> RDF as a graph model is really to support data aggregration through
CG> graph matching -- in the way that TAP does for example
CG> (http://tap.stanford.edu/). You shouldn't lose sight of the fact that
CG> the Semantic Web movement if actually about DATA INTEGRATION and not
CG> about AI at all. This sometimes gets forgotten :-)
that's true. I sometime try to "de-worry" people about the complexity
of some languages and then make some (only slightly) simplifying
statements such as "RDF is just labeled graphs or triples -- done".
Please forgive =B-)
CG> e) One of the import things you should not lose sight of is the
CG> "semantic continuum" (trademark Mike Uschold). By this I mean that some
CG> of your ontologies will be no more that simple hierarchies, and some
CG> will need subsumption reasoning. And some will want other forms of
CG> reasoning that take them out of the decidability camp. Moreover, you
CG> will want to slide along this continuum --starting simple and then
CG> selectively elaborating. Or even going the other way and selectively
CG> simplifying. Have a language that allows you to say as little or as much
CG> as you want in the same framework, and provides reasoning support for
CG> evolution, is darn handy. That is why we use DAML+OIL.
CG> f) Finally -- take care about ontology development vs ontology
CG> deployment. You could use all sort of fancy stuff in design (and I would
CG> argue that it is needed - see (a) ) but then generate a simple
CG> representation for deployment -- say in RDF, or even Topic Maps. You
CG> cannot assume in the Grid or the Web that your deductive database or
CG> your reasoning engine is available, but you still want to get something
CG> out of your metadata.
CG> Conclusion -- there is no one technology, we need a basket of
CG> technologies working in consort that do the best job for what you need.
I couldn't agree more -- and I hope that if you take the conjunction
of my statements that you can derive "false" (i.e., derive an
More information about the semantics