ontolog-forum
[Top] [All Lists]

Re: [ontolog-forum] Probabilistic Ontologies

To: "[ontolog-forum] " <ontolog-forum@xxxxxxxxxxxxxxxx>
From: Kathryn Blackmond Laskey <klaskey@xxxxxxx>
Date: Tue, 12 Jun 2007 19:32:11 -0400
Message-id: <p0611046fc294dd690b82@[192.168.0.108]>
Pat,    (01)

>Can you point to where I can find out more?    (02)

http://www.pr-owl.org is the web page my ex-student Paulo Costa 
created for PR-OWL, the probabilistic ontology language.  PR-OWL is 
an OWL upper ontology in which you can augment ontologies with 
probabilities using a first-order probabilistic logic called MEBN 
(Multi-Entity Bayesian Networks).  MEBN is one of a generation of 
first-order probabilistic languages that were developed in the late 
90's and early 00's.  Basically, these languages can be viewed as 
ways to specify a probability distribution over interpretations of a 
first-order theory. We (I and students and colleagues too numerous to 
mention but without which this work would never have happened) 
started working on MEBN in the mid 90's while trying to build 
probability models for DARPA, after realizing we needed greater 
expressiveness than Bayesian networks could offer.  For the theory, 
you can read
   http://ite.gmu.edu/~klaskey/papers/Laskey_MEBN_Logic.pdf,
which is currently under revision for AIJ. (Except that I spend too 
darned much time writing to this forum and need to spend more cycles 
finishing up the revisions!)    (03)

A student at University of Brasilia is developing an open-source MEBN 
reasoner, which saves in PR-OWL format and has a GUI that allows you 
to enter graphical probability models.    (04)

For the past two years, I have co-organized workshops at the 
International Semantic Web Conference on uncertainty.  See
   http://sunsite.informatik.rwth-aachen.de/Publications/CEUR-WS//Vol-218/  and
   http://sunsite.informatik.rwth-aachen.de/Publications/CEUR-WS//Vol-173/    (05)

Another workshop is planned this year in Korea, although I'm not 
organizing it because I'm co-chairing a W3C Experimental Group on 
Uncertainty Reasoning for the World Wide Web
   http://www.w3.org/2005/Incubator/urw3/.
We are developing a set of use cases for uncertainty in the World 
Wide Web.  The ones we've done so far are pretty sketchy, but they 
can be found on our wiki at
   http://www.w3.org/2005/Incubator/urw3/wiki/FrontPage.
Only XG members can modify the wiki, but anyone can read them and 
comment to the public email list.  Anyone who works for a W3C member 
can join; if you are interested and don't work for a member, you can 
ask to join as an invited expert.  We started in March and our 
charter lasts for a year, so it wouldn't be too late to join.    (06)

The wiki also includes a sketchy reference list; we know it's 
incomplete, and would be glad to be told of things we've left off.    (07)

>>  (The state of usability of the tools is
>>another matter!)  My colleagues, students and I have built several
>>toy probabilistic  ontologies.
>
>I want to see them !!    (08)

The Star Trek ontology from Paulo's dissertation can be downloaded 
from http://www.pr-owl.org.  You can open it in Protege. It imports 
the PR-OWL upper ontology. It's explained in Paulo's dissertation, 
which is also available at pr-owl.org.    (09)

Several people are using PR-OWL to develop ontologies; I'm consulting 
on some of the efforts.    (010)

There is a group at University of Maryland that has developed a 
language called Bayes-OWL that enables you to specify a Bayesian 
network in OWL.  Bayes-OWL uses standard Bayesian networks; hence, 
isn't sufficiently expressive for the full range of ontology 
development needs.    (011)

Related work is KEEPER, developed by IET, Inc. IET (where a number of 
my major partners-in-crime currently or formerly reside) has a rich 
probabilistic modeling language called Quiddity*Suite, that allows 
you to put probabilities on the values that fill a slot in a frame, 
including probabilities on reference slots. The probability 
distribution for a slot can depend on the values of other slots in 
the same frame, or on slots of related frames (e.g., the probability 
that I'll get an A in a course can depend on the teaching ability of 
the teacher of the course).  This results in a very expressive 
probabilistic language. You can make instances of frames, specify 
values for some of their slots, and use Bayesian inference to obtain 
probabilities of the slots whose values aren't specified.  KEEPER 
allows you to build an OWL ontology that can be translated into 
Quiddity frames.  Quiddity*Suite is being integrated with the Netica 
Bayesian network package.   See http://www.norsys.com.    (012)

>>  I've worked with colleagues to build
>>artifacts that were called "probability models" but had features that
>>one associates with ontologies, i.e, were built in expressive
>>probabilistic languages with have types and individuals, subtypes and
>>inheritance, attributes and relations.  I wouldn't call any of these
>>"well-engineered probabilistic ontologies" for the reason that they
>>were built for clients to solve specific problems, and had to make
>>compromises because there were no good upper ontologies for some of
>>the things the system had to do, and there was no time to build them,
>>so hacks were devised that worked for the purpose of the specific
>>problem but would give a principled ontologist a queasy feeling.
>
>Compared to what is out there, Im sure they will be models of 
>clarity. Seriously, is it possible to see what kind of thing you 
>have been doing along these lines?  Offline and informally, of 
>course, if you prefer. I am SERIOUSLY interested in understanding 
>this.    (013)

There are lots of papers on my web site that give an idea of the kind 
of thing I've been doing.  My web site is horribly disorganized -- 
people keep telling me I should make some kind of organizational 
scheme instead of just a long list of papers -- but that would take 
time I'd rather spend writing emails to Ontolog Forum.  :-)
   * http://ite.gmu.edu/~klaskey/papers/Wright_Laskey_Credibility.pdf 
describes what could be called an oversimplified start at an ontology 
for source credibility.  It's based on Dave Schum's extensive 
research on modeling credibility of sources.  My former student Ed 
Wright developed these ideas further into a demo that shows how you 
can fuse evidence from multiple sources, incorporating source 
credibility, to draw interesting conclusions about naval scenarios. 
ONR declined to fund a proposal for follow-on work that included, 
among other things, turning this into a real ontology.
   * http://ite.gmu.edu/~klaskey/papers/BRIMS04_InsiderThreat.pdf 
describes a project devoted to detecting insider threats in 
information systems -- specifically people in highly secure 
environments who are looking at documents they are cleared to see but 
aren't related to what they are "supposed" to be doing.  (I know: you 
wouldn't want somebody looking over your shoulder like that, but it's 
part of the deal you sign up for when you work in a highly secure 
environment.)  We built an organization and task ontology 
(non-probabilistic), a system usage ontology (non-probabilistic) and 
an insider behavior ontology (probabilistic).  These mediated the 
interaction between a system that captured user document accesses, a 
data mining system (one student's doctoral dissertation) that 
categorized the relevance of documents to topics, and the 
probabilistic reasoner -- Quiddity*Suite -- that assessed the 
likelihood that a user was a threat (another student's doctoral 
thesis).  ARDA pulled our funding two months before we were going to 
conduct an experiment in which we would have given a bunch of 
students a research task, given half of them a surreptitious research 
task, required both groups to write reports on their "real" task and 
also required the second group to write a clandestine report on their 
surreptitious task, and performed a blind test to assess our ability 
to catch the students doing the clandestine task.  ARDA had major 
funding cuts that year, and cut everything that they didn't see as 
just about ready for deployment.  We were far too researchy.
   * http://ite.gmu.edu/~klaskey/papers/LaskeyLevitt_4708-30.pdf 
describes a probability model (wouldn't call it an ontology, but it 
could be extended to one) that models a bioterrorist anthrax attack. 
We did this one for fun.  It was unfunded.
   * http://ite.gmu.edu/~klaskey/papers/Levitt_Laskey_Inf_JP.pdf is a 
pre-publication version of a paper later published in the Cardozo Law 
Review.  It describes a probability model of murders, again that 
could be extended into a bona fide ontology, applied to a murder case 
that actually occurred in France. (Now you see what I meant above 
when I said partners-in-crime -- my co-author is the CEO of IET.)  To 
be truthful, the model was engineered to be applied to this 
particular murder case -- that's why I would get queasy calling it an 
ontology. But we worked hard at abstracting it into a generic model. 
This paper was also done for fun.  (Ask my family about the nights I 
was up during the wee hours during our beach vacation.)    (014)

This is what I been doing with a large chunk of my time for the past 
n years -- engineering probability models to specific problems, but 
trying to develop them so they will be extendable, and also trying to 
develop the necessary theory to do this right.  In my experience, 
funding agencies balk at the labor-intensive task of building out 
real ontologies, especially in today's climate where every dollar of 
government funding is going to things that can be deployed tomorrow. 
I agree that the job is tedious and labor intensive -- but that means 
we need to fund the basic research that's necessary to make it less 
tedious!  This seems to be a very hard sell.    (015)

>>I expect this situation to dramatically change over the next decade.
>
>Oh come on, in a decade I might be dead.    (016)

I hope not!  The world would be a lot less fun without your crochety 
emails to look forward to!  :-)    (017)

>I want to know about it NOW :-)    (018)

How's this for starters?    (019)

Kathy    (020)

_________________________________________________________________
Message Archives: http://ontolog.cim3.net/forum/ontolog-forum/  
Subscribe/Config: http://ontolog.cim3.net/mailman/listinfo/ontolog-forum/  
Unsubscribe: mailto:ontolog-forum-leave@xxxxxxxxxxxxxxxx
Shared Files: http://ontolog.cim3.net/file/
Community Wiki: http://ontolog.cim3.net/wiki/ 
To Post: mailto:ontolog-forum@xxxxxxxxxxxxxxxx    (021)

<Prev in Thread] Current Thread [Next in Thread>