controlled NLs

["Rich Cooper" <richcooper@mindspring.com>]

I've finally finished reading Luc Steel's huge number
of papers, and I really like his approach, though his
example systems seem very small in extent.  Menno
van Zaanen's papers and thesis on "alignment based
learning", together with Steel's ingenious grounded,
distributed, communicating multiagent systems seems
like the kind of project that could finally bring some
degree of realism to nlp question-answering systems,
like simplistic agents in the appliances.

Has anyone published work about using controlled
nls in this kind of environment?

It occurs to me that a primitive-based controlled nl,
using something like Wierzbicka's NSM primitives
bound to specific synsets would be the kernel forms
of a viable controlled nl.  That would be the grounded
part.  Steel's multiagent approach would provide a
self generating set of hypotheses about relating phrases
to Wierzbicka's kernel primitives, learning new
nonterminal concepts from interaction with a teacher.
Finally, van Zaanen's approach of learning structure
from phrases that have recognizable NSM words
could build up from the grounded stage, one layer
at a time.

Of course, this wouldn't cover very much of English,
but it might be a start at producing better communicating
agents than we presently have in our appliances.

One issue would be to get representative teachers
to spend time with it.  It seems that the Cyc approach
didn't organize its small number of "teachers" on a
grounded basis, and that may be why Cyc never quite
got to the critical stage.  A learning nlp q&a should
do its own internal organization instead of having that
choice thrust upon it IMHO.

Any thoughts on these issues would be appreciated.

Sincerely,
Rich