Re: Practical knowledge representation
On Monday 04 April 2005 23:31, Rich Cooper wrote:
> Rob Freeman wrote
>
> > Something I've wanted to post here for several days, talking of practical
> > approaches to knowledge representation, who here is aware of Ted
> > Pedersen's "Senseclusters" project on SourceForge?
> >
> > http://www.d.umn.edu/~tpederse/senseclusters.html
> >
> > For anyone wanting to play around with open source software for knowledge
> > representation (Rich?) this strikes me as about as useful a study as any.
> >
> > -Rob
>
> Thanks Rob, I've subscribed to his lists, and I'll start reading his
> papers. This does look like a good lead for my own work.
Great.
I don't think Pedersen fully sees the potential for defining meaning to match
each new problem. Like everyone else he and his students are focused on
(fundamentally ill-defined?) industry standard "classification" tasks like
SenseEval. But I think their representation could be used in that way (to
implement meaning as the _relevant_ organization of information, not to find
a canonical organization) if once people realized that is what they need to
do.
As a sketch of one kind of drug conflict application you might try to build
along those lines (relevant organization of information, instead of canonical
organization):
If you can identify a class of drugs (say drugs which slow the heart) by their
clinical contexts (just cluster descriptions in texts), then you should be
able to identify a description of those drugs (text actually saying they are
"drugs which slow the heart") by those same contexts. So potentially
presenting the system with a list of drugs which slow the heart could cause
the system to isolate the cluster, produce the description "drugs which slow
the heart", and warn of possible compound effects.
You would just be using the similar drugs (similar in some way possibly unique
to that particular collection) to produce a cluster which relates them, and
prompt a description which is also unique to that cluster.
That would be an application to warn of compound effects. I expect you would
be able to mine clusters to find other types of conflicts, too.
The nice thing is that with information represented in unstructured
distributions of associations, you have the flexibility to find such new
clusters (new "meaning") and thus potentially new conflicts.
-Rob