I find the practical issues in applying a single
architecture
to a wide ranging enterprise to be awesomely
consuming.
As a programmer, I would find another project
rather than
get into that deep a mess.
Looking at the difficulties we've had in SUO coming
to
agree on any kind of ontology, how could an
enterprise
indoctrinate every programmer on the concepts s/he
is using
so that they become formally equivalent?
I'm beginning to think that starting with an
ontology
instead of ending with one is the wrong
approach. Perhaps
the ontology should be generated from the
subordinate
systems using some kind of analysis tools.
John, your reference to Luc Steel's work has
been keeping
me very busy reading his papers. I'm not done
yet, but here's
some quotes from one of them:
"The approach we have used for lexicon formation is
different
from the more traditional Quinean approach.
Quine assumes
that agents learn the meaning of words by making
progressive
inductive abstraction from the situations in which
they observe
a particular word-object relation. ... In
contrast, we have adopted
a Wittgensteinian approach, where agents invent
words and
meanings as part of a language game, formulate
different
hypotheses about the meanings of words used by
others and
test these meanings in their own language
production. The
evolution towards lexicon coherence in
the population (where
one word has one dominant meaning and
one meaning has
one dominant word) is a collective
phenonmenon triggered
by responses of the system to new
situations in which the
multiple meanings of a word are no
longer compatible with
each other."
...
"The most valuable result of our experiments is
that we are
able to demonstrate a scale up. New meanings
arise when
the environment becomes more complex and the
lexicon
keeps adapting and expanding to sustain
successful
communication."
...
"In the literature, intelligent generalisation,
specialisation, or
elimination operators are often ascribed to the
individuals
acquiring language. This is not the case in
the present
paper. Agents have very minimal forms of
intelligence.
We have observed that an agent sometimes starts to
prefer
a more general category for a word (or a more
specific one),
but this is not due to an explicit generalisation
operation in
the agent, it takes place as a side effect of the
repair action
undertaken after a failing game."
In my own experience, we learn language easily as
children,
but have to be taught about generalization,
specialization,
and other abstractions of language after we have
already
learned to communicate. So there must be some
kernel
of truth in Steel's point of view.
For ontologies, that means they appear after each
subordinate
system has been gotten working properly in its own
domain.
But if that's the case, then generating an ontology
from the
component systems may or may not even be
helpful.
JMHO,
Rich