SUO: RE: Re: Missing Ingredients
John F. Sowa wrote:
> Richard Cooper wrote:
>
> RC> John, can you be more explicit about how we can 'directly
> > address context'?
>
> I consider contexts to be as large or as small as you like.
>
> In conceptual graphs, I represent a context by a concept box
> that contains a nested CG. The context consists of anything
> that may be described by a CG. Since a CG can be arbitrarily
> large or small and it can describe large things and small
> things (by any literal or metaphoric definition of size),
> almost anything could be a context.
>
> If you would prefer English to CGs, I would say that a context
> may be anything that you can describe by an English sentence
> of any size. (And I would permit as many clauses as you like
> to be connected by semicolons.)
>
> John
I like that as a rough approximation. I'm thinking in terms
of a question-answering system equipped with an initial
ontology. I define: there is a context C[k] for the Kth sentence
which describes just what is in that sentence itself. Also,
there is a context C[*] for the preceding k-1 sentences that has
developed from the first sentence in the session. So context
is a running database.
Given that scenario, context should designate the individuals
(by which I mean objects of interest) that are being tracked
by the computer's side of the conversation.
To be a little more specific, consider WordNet (or Cyc, or
others) as a basis for the QA system. Presently, neither
of these comes with a software envelope to process the axioms
and the data developed during the conversation. That envelope,
in the scenario I'm considering, should be the QA system, and
some of the results of that QA system should be:
- improved feedback from operators as to the accuracy
and validity of the present ontology,
- improved world knowledge,
- measure of effectiveness of the session.
This is the minimum set of requirements for a system that
can provide a useful service while growing its knowledge
base through situated interaction.
The context is either committed (if the MOE is adequate) or
deleted (if the MOE is high enough).
Of course, there need to be other things, such as "normal"
behavior scripts, various "emotional" drive simulations, and
other mechanisms that mimic linguistic behaviors on the
part of the QA system.
Does that sound like a reasonable (though difficult) project
to build an ontology from multiple users through conversational
interaction?
Sincerely,
Rich