[From Bill Powers (921025.1900)]
Avery Andrews (921026.0936) --
So far so good: the higher system controls for meaning by finding
A lower-level sentence that produces that meaning. Reorganization
takes place when no sentence with a suitable meaning can be found.
I don't think we're talking about quite the same thing here. The
higher level system here is supposed to work by controlling for a
perception of finding (appropriate) meanings in what people are
saying - no sentences are getting produced here.
"Finding" a meaning means perceiving one, doesn't it? And perceived
meanings are found from perceived sentences, aren't they?
I'm really fumbling my way through this, because there are so many
relationships that I don't understand. Here's a rough picture of what
I'm visualizing -- see if it fits your ideas, or could be made to do
so.
desired
meaning (nonverbal perception)
>
perceived meaning-->COMP-------error --->
> >
> >
input function output function
> > > > > > > >
sentences sentence-variations
> >
<---Imagination or real words <-----
Going up the left side, you perceive sentences, which then give rise
through some sort of input function to perceived meanings. The
perceived meanings (which are not in words, but in terms of
perceptions of the things that the words mean) are compared directly
with the desired meanings. The error acts through an output function
to vary the construction of sentences. These are the sentences that
are perceived, closing the loop. Of course when you close the loop by
actually uttering the sentences and hearing them, other people hear
them too, which takes care of producing utterances.
I would assume that our equipment for perceiving sentences and
meanings become organized at least in important ways by learning to
create them.
There are clearly more questions than answers in this arrangement. I'm
just trying to get across the general PCT concept that we control
perceptions, inputs not outputs, by VARYING actions. An error in
meaning results in a correction of sentence structure such that the
new sentence is perceived and gives rise to a new perceived meaning,
which one hopes is closer to the desired meaning. Instead of seeing
the process of converting meanings to sentences as an output process,
we see it as a process of controlling inputs, where sentence structure
and meaning are both perceptions, not outputs.
To make a viable model out of this, it's necessary to specify the
kinds of process that have to go on in the input and output functions.
The input function must involve conversion from symbol structures into
nonverbal experiences, so memory and association would be needed. The
meaning-errors would have to be specified in terms of the dimensions
of the meaning perceptions -- this is really a schematic of many
control systems acting in parallel, each concerned with a different
dimension of meaning. Then we have the problem of how to convert a
difference between an intended meaning and a perceived meaning (in one
of the dimensions) into an appropriate kind of change in the sentence
-- that is, a change that will bring the perceived meaning of the new
sentence closer to the reference-meaning.
I think the same questions are really implicit in a command-driven
model without the feedback loop -- except that the conversions have to
go in the other direction. Something has to convert an intended
meaning into a sentence that has that meaning. Once you see the
feedback way of doing this, the other way (generative) shows its
holes, doesn't it? And without the feedback loop, generation of
meaningful sentences has to be treated separately from perception of
the meanings in given sentences. The feedback model uses the same
perceptual system in either case.
I'm thinking of the production system as a sort of transducer that
turns 'meanings', some sort of internal structure, into an overt
performance.
That's always seemed to me to be the weakness in the top-down kind of
model. Given a meaning to express, there are very large numbers of
sentences that will convey it. What kind of transducer can convert
from a single signal into "the right" output among many outputs --
without any feedback to check if that was indeed the right output?
This is the same weakness that exists in the general traditional top-
down concept of the brain, where high-level systems issue general
commands, which then -- somehow -- get turned into just the muscle
tensions that produce just the consequences that could be viewed as
appropriate executions of those commands. How do the lower-level
systems know that what they are doing, in the current environment,
will turn out to be a valid instance of the commanded result? What
tells them that moving to the right will execute the command "mail the
letter?" The command, being general, can't contain that information.
Where does the information come from? This "final common pathway"
concept is just not bloody likely.
The control-system model completely takes care of this problem. It
removes all the ambiguities in a top-down command-driven model.
What is varied first is the sentence; that (going upward in the
perceptual systems) varies the meaning; that creates an error that
motivates a change (going downward now) in the perceived sentence.
This process continues until the perceived meaning matches the desired
meaning. Any sentence that does in fact produce the desired meaning is
"correct". The only reason it might be changed more is that there is
another meaning-control system monitoring for some other aspect of
meaning evoked by the same sentence. Then the process of varying the
sentence will continue until both meanings are perceived.
When you're just passively listening, the sentences evoke the meanings
in the upgoing paths, without the errors (if any) leading to
corrective action. Or maybe the kind of corrective action changes --
you say "Do you really mean I should turn right, into the river?"
There are obviously some problems in working out a plausible instance
of such an arrangement. How does an error get turned into the right
kind of change in a sentence? It doesn't have to produce the correct
sentence in one jump, but it has to alter the sentence in a direction
that lessens the error. I can think of some sorts of errors, like
wrong person, wrong order of events, wrong outcome, incorrect
adjective or adverb, mistaken referent, and so on. These would be the
dimensions in which the meaning can change, and I would assume that we
learn how to vary meanings by altering sentence structure and word
choice; that would convert directly into making errors get smaller by
varying sentence structure.
I can see where exploring this model would require some novel
experiments with language and meaning. It would be informative to see
how people correct sentences that convey wrong meanings to them (even
if the sentences are grammatically constructed). For a particular kind
of meaning error, what sorts of changes in sentences do people make in
order to correct them? This would begin to show the dimensions of
meaning-error signals, and the dimensions along which sentences can be
varied to alter particular kinds of meanings.
This takes a completely new approach to understanding how language and
meaning hook together, I suppose. I'm not the person to do it. But
whoever does do it is clearly going to have to break with old habits
of analysis. The first habit to go would have to be the idea that
meanings produce sentences produce utterances. In PCT the chain goes
in the other direction... it's all perception.
Perhaps we should pause to get this ironed out before proceeding.
That may be somewhat optimistic, but we can try.
P.S. Given three nouns A, B, and C, you can create the sentences
The A B'd the C, and
The B A'd the C.
The captain hogged the ice-cream, and
The hog captained the ice-cream.
The only way to judge that one is allowable in ordinary discourse and
the other is not is through the meanings of the words. That can't have
anything to do with language itself, because the fact that hogs don't
captain ice-cream is an accident of experience; given other
experiences, we might see immediate sense in the second sentence but
not the first, depending on what perceptions "hog" and "captain" are
attached to, and what we can imagine hogs and captains doing. There's
something about the way words are presented that connects to the way
meanings are evoked in juxtaposition and sequence -- that's the input
function above. But once the meanings are evoked, the rules that apply
aren't those of language, but of perceived reality in general. We
don't judge that hogs can't captain things because of language, but
because of our experiences with hogs and captains.
That must sound like utter nonsense.
···
-------------------------------------------------------------------
Best,
Uncle Bill P.