Re: principles of collective organization
Hi Fred,
I'll reply to some of your detailed points re. cognitive primitives
parameterizing language etc. off-list. If anyone wants a copy let me know.
On Thursday 12 January 2006 00:00, Frederick B. Kintanar wrote:
>
> When you say collective aspect of language, do you mean "all parts of
> the brain" or do you mean "social" or what?
I mean grammar and meaning modeled as distributions of word associations.
You can model meaning as collections of examples, grammar also. You can
link the two.
> > This is simply a thing which is not being done. Even so-called
> > distributional methods are universally, to my knowledge, limited by the
> > expectation that there are underlying primitives (or at least primitives
> > by consensus, c.f. Luc Steels at Sony, etc.)
>
> I am not familiar with distributional methods, a pointer would be helpful.
How much do you want? You can split it into two halves: before Chomsky and
post-Chomsky. American Structuralism was essentially _all_ distributional.
Chomsky killed it in the '50s by pointing out it led to inconsistent
representations (=no primitives). Now that people have forgotten Chomsky's
objection distributional methods have become a popular machine learning
technique, but they are searching for primitives again.
e.g. Schuetze, "Dimensions of Meaning", in Proceedings of Supercomputing(?),
1992, http://www2.parc.com/istl/groups/qca/papers/Schuetze-Meaning92.ps.gz.
Dominic Widdows' new book is a recent publication which I believe to be in
this tradition ("Geometry and Meaning",
http://infomap.stanford.edu/book/chapters/chapter7.html). Interestingly
Widdows does draw an analogy between lexical semantics and Quantum Mechanics,
on some level.
The agent-based stuff is slightly different.
From memory Luc Steels (www.csl.sony.fr) takes populations of agents and shows
they agree on symbols to achieve their goals. It is a form of emergent
meaning. A researcher Timo Honkela in Finland wrote to me a year or two back
and seemed to be doing something similar:
http://www.cis.hut.fi/~tho/publications/honkelawinter2003tr.pdf
There are sure to be others. They had a conference
(http://www.cis.hut.fi/AKRR05/).
The agent-based work distinguishes itself by being emergent, but really it is
still limited by the expectation of a primitive, a "consensual" primitive in
this case. What they find is emergent vocabulary.
Then there are the distributional classifiers like neural-networks.
Case-based reasoning, MBL, AML, have an edge on neural-networks by not having
any learned representation. That gives them more descriptive power, but they
are parameterized by their solution set, another kind of primitive, so it
comes down to much the same thing.
> >...to my knowledge, a true re-evaluation of cognition and language in
> >terms of ad-hoc structure according to "principles of collective
> >organization" is just not being done.
>
> I'd like to suggest that Channel Theory and some other compatible
> principles provide a good starting point. Working with IFF also gives a
> clearer reason why such a program of work should be done within the SUO
> group.
This is OK, but it limits you to categories which already exist. Probably
categories which have been laboriously elaborated at some great cost by hand.
At some point people won't want to develop systems of categories by hand any
more. Then you will need some way to find them automatically.
I suggest that distributional methods are fine. You just need to reorientate
them away from the mistaken goal of finding primitives, and tune them to
search for "relevant" structure instead.
I think Internet search engines already work that way to a large extent by
default. I suggest the techniques they use for generalizing information now,
essentially naive inverted lexical indexes and page-rank, might be
considerably improved with distributional generalizations based on those used
by language (to model meaning.)
-Rob