LINGUIST List 9.544

Wed Apr 8 1998

Disc: NLP and Syntax

Editor for this issue: Martin Jacobsen <martylinguistlist.org>


Directory

  1. Dan Maxwell, NLP and Syntax

Message 1: NLP and Syntax

Date: Sun, 5 Apr 1998 12:27:04 -0400
From: Dan Maxwell <100101.2276compuserve.com>
Subject: NLP and Syntax

Peter Menzel claims that neural networks(NN's) are a vast
oversimplification. Let me observe some ways in which this appears to
be true at present. First of all, nobody's NN's, as far as I know,
incorporate all the different kinds of knowledge that real life
language use involves.
 
My own NN's deal mostly with morphology and syntax; others deal with
phonology; people in artficial intelligence deal mostly with other
kinds of real world knowledge. Someday we'll have to try to tie all
of these things together. Second, NNs are represented in
two-dimensional space, since that is what paper is best suited for.
But our brains are three-dimensional; so maybe the ultimate NN should
look more like the representations of the solar system that we find in
planetariums. Third, in the best generative grammar tradition, nobody
distinguishes speaker and hearer functions, although, as I think Peter
Menzel mentioned earlier, these do seem to be more or less distinct in
the brain (Broca's area vs. Wernicke's area). But this too can be
worked in someday.

Whatever force these and other possible objections might have, I think
NN's are interesting as a way of forming explicit hypotheses about the
relationship between language and the brain. The way to find out what
is wrong with any proposal is to make it as explicit as possible, test
it, pinpoint its failings and revise it accordingly.

I still don't know what the usual sense of the word "algorithm" is
that Peter Menzel is referring to, but it is true that special areas
tend to use words in special ways that are not yet in the
dictionaries. Maybe this is what has happened here.
 
I'm not sure how to interpret the claim that a formailistic framework
which is suitable for a computer is not likely to be indicative of how
language works in the brain. On the one hand, it is clear that we
won't find information like NOUN spelled out explicitly. The best we
can hope for is some configuration of neurons or other gray matter
that encodes such information. We can draw on the distinction made
earlier by Peter Menzel between psychological behanvior and
neurological behavior. NOUN is information at the psychological
level; if we are correct in positing such information in our grammars,
then we have to suppose that this is somehow translated into
neurological behavior. Similarly, when such information is posited in
computer programs, it has to be translated into some series of yes-no
decisions that the machine can execute.

Maybe the objection is to formalism. But formalism is just a
convenient way of representing things we could express in words. Once
we master the formalism, then it is actually easier to work with than
words. Presumably everybody in this discussion mastered the formalism
in the expression: "2 + 2 = 4" back in grade school and found it
useful in many ways that had nothing to do with computers. The
formalisms that we use in linguistics are a little harder than this,
but we try to stay open to proposals about easier ones. And our
linguistic formalisms are useful for computer programmers who have to
transform them into computer languages, which at the highest level are
beginning to resemble English. Dan Maxwell
Mail to author|Respond to list|Read more issues|LINGUIST home page|Top of issue