Language as process

[From Bill Powers (931019.1100 MDT)]

Martin Taylor (931018.1700) --

I don't know, but it seems to me that we are getting more and
more confused over "contrast" "phonemes" "words" "morphemes"
"phonetics" and the like.

Very nice summing up, but you left something out: LAP, or
language as process.

Your analysis focuses primarily on language as an emergent
external-view phenomenon (language as artifact) and language as
evidence of high-level control (language in use). I agree to both
of these distinctions. But, for language in use, saying THAT we
interact with others and mutually reorganize or WHY we do so does
not answer the question of HOW we do this or WHAT organization in
each individual is the result of all this.

The question of HOW is what concerns me as a modeler. The
appearances of language as artifact are seen without the
discipline of a model; there is no constraint that makes the
description correspond with a description of an underlying model.
Thus the objections that Rick and I have had to a statement that
people perceive contrasts. When you ask HOW you would build a
perceptual function that would perceive contrasts between words,
you find that you have to make complicated assumptions. Even when
you do make such assumptions, the result is not perception of
contrast per se, as a dimension of experience, but an effect on
perceptions which are not themselves contrasts, but phonemes (or
whatever). Contrast is an outcome of these effects, not their
explanation.

If you truly built a contrast-perceiver, it would report its
world in terms of a degree of contrast: 6 units of contrast, 25
units of contrast, 0 units of contrast. That's all the signal
would indicate: contrast. It would not, could not, also indicate
contrast of WHAT with WHAT. Any two words that differed by 6
units of contrast would be perceived as the same contrast: 6
units. The contrast-perceiver could not tell which words the
contrast occurred between, or what features of the words account
for the sense of contrast-ness. This is entirely appropriate for
a high-level perception; the details are lost, or rather the
details remain at the lower levels that handle details.

This problem keeps recurring in our linguistics discussions, as
well as others. I've described it as the difference between
CHARACTERIZING a model and DESCRIBING the model. When you
characterize something, you say something about what the model
accomplishes, or offer a metaphor for its behavior, or give some
other general assessment of the model that FAILS TO MENTION WHAT
THE MODEL ITSELF IS. So you can say that the stretch reflex is a
system that corrects errors, or that makes its input match its
reference level, or that is purposive in some regard --
statements which may be true, but which leave us in ignorance of
what the stretch reflex IS in terms of functions and signals in
the neuro-muscular system.

So it is with a concept like contrast. To say that phonemes are
identified by their contrast with other phonemes really just
characterizes the recognition process in a metaphorical way. It
is as though something were looking at the difference between
phonemes and using this difference as the basis for
identification. That may be an entirely reasonable metaphor, but
it takes us no closer to being able to say how phonemes are
distinguished.

In fact, it begs the very question that we try to answer by
modeling: what it is that detects a difference between two
phonemes so that a judgment of contrast or no contrast can be
made? Before any difference can be noted, there must BE a
difference -- that is, the individual phonemes (or whatever the
units are) must be represented differently, individually. If
their representations are identical, there is no magical way that
a difference between them can be detected.

When we understand how these differences are generated, we will
be able to solve the apparent paradox of identical perceptions
being treated differently and different perceptions being treated
identically. We will see that if there is a different treatment,
there is an actual difference, and if there is no different
treatment there is no actual difference. We will find that the
paradox existed only because we were assuming an inadequate model
of how the perceptions in question work. Perhaps in defining a
phoneme we have unwittingly created the paradox. In that case we
have to find a more useful unit of organization that doesn't lead
to mysteries. Perhaps the truth is that the basic unit of
recognition is not a single phoneme, but a space-time pattern of
phonemes. That's the assumption I've been using. With a space-
time pattern, there's no need for any concept of contrast as a
causal entity. The OUTCOME of perceptual processes based on such
patterns may well be a judgment of contrast, but contrast is not
the basis for the judgment.

Well, I'm probably still butting my head against the same stone
wall.

ยทยทยท

-------------------------------------------------------------
Best,

Bill P.

[Martin Taylor 931019 17:45]
(Bill Powers 931019.1100)

Martin Taylor (931018.1700) --

I don't know, but it seems to me that we are getting more and
more confused over "contrast" "phonemes" "words" "morphemes"
"phonetics" and the like.

Very nice summing up, but you left something out: LAP, or
language as process.

Good point. One I had entirely ignored.

Your analysis focuses primarily on language as an emergent
external-view phenomenon (language as artifact) and language as
evidence of high-level control (language in use). I agree to both
of these distinctions. But, for language in use, saying THAT we
interact with others and mutually reorganize or WHY we do so does
not answer the question of HOW we do this or WHAT organization in
each individual is the result of all this.

Yes, you are right. To some extent, the Layered Protocol theory
concerns itself with this aspect. I did not describe it in the posting,
and in any case I suspect you would find it too vague to consitute
a working model. We are working on that, trying to build the tools
that would allow us to implement it in a computer interface. The theory,
at present, says nothing about how the various protocol nodes come to
be organized. At present, I assume that processes general to PCT
reorganization apply equally to the LP structure. (A protocol node
in LP is analogous to, and I hope will prove identical to, an ECS in
a standard PCT hierarchy).

The question of HOW is what concerns me as a modeler. The
appearances of language as artifact are seen without the
discipline of a model; there is no constraint that makes the
description correspond with a description of an underlying model.
Thus the objections that Rick and I have had to a statement that
people perceive contrasts.

...

If you truly built a contrast-perceiver, it would report its
world in terms of a degree of contrast: 6 units of contrast, 25
units of contrast, 0 units of contrast. That's all the signal
would indicate: contrast.

It is against this view of contrast as a continuous variable that I
have been arguing. I work from a viewpoint, often expressed with
various degrees of force, that categories and above do not live in
a continuum perceptual space. The sensory signals input to their
PIFs are continuous--they are the perceptual signals of lower levels.
But at and above the category level, there is no topological neighbourhood,
no continuity. You can't get 6 units of contrast. You can get "same"
or "different." The degree of difference is something that applies
to the contributing lower level perceptions. "Different" applies between
two categories that contrast. "6-units different" applies to the complex
of sensory inputs to that categorical perception.

Let me continue with Rick Marken (931019.1200)

Martin Taylor (undated -- but about 931013)

It was intended to be sent to Rick privately, but I did send it to the
net. Sorry.

I don't think you get what Bruce is trying to say. If I read him right,
one circumstance he is trying to cover is that the SAME phonetic signal
is perceived as a DIFFERENT phoneme in different situation or verbal
contexts. That in no way casts aspersions on the correctness of PCT.

It is an incorrect representation of the PCT model of perception. A
perceptual signal (in the model) IS a perception. So a signal whose
magnitude represents, say, the degree of "d"-ness IS the perception of
"d"-ness.

I take this as a claim that there can be no level of control for which
the perceptual signal has a discontinuous set of values. I have been
trying to dispute this claim. If to do so is to propose an incorrect
representation of the PCT model of perception, so be it. I would then
dispute the word "the" in "the PCT." There is no dispute as to whether
"A perceptual signal (in the model) IS a perception." So it is.

So called "context effects" would be included in the
perceptual functions at the next higher level, the ones that get
the "d"-ness signal as one of their inputs.

I do think I discussed that somewhere, recently.

Bruce seems to think
of contrast as a perceptual variable that is controlled during
speech; this sounds wrong to me so I just want to see Bruce's
model of contrast (speech) perception.

I think the same. I think I explained how it might work in the same
"somewhere" posting, or one near it. I don't want to do it again, at least
not right away. It involves the perceptions at the next higher level,
and includes the perception of the partner's state.

Given the way Bruce (and linguists in general, apparently) talk about
contrast I would imagine that they would describe "straight lines
connected at the top" as a "contrast" in capital lettering (because
it distinguishes "A" from "H"). As I recall, this is the way
"distinctive feature" models of speech classified phonemes;.

No, it would be the A and H that would contrast. If "connected at the
top" were a category, it might appear in some contrasts, usefully with
one such as "disconnected at the top." But more probably, it would be only
a lower-level perception the magnitude of which would depend on how closely
the lines touched at the top without crossing. It would depend on the
person perceiving the lines.

In distinctive feature models, phonemes are located in a multidimensional
space where there are only two values (maybe three sometimes) on
each dimension (distinctive features). There are dimensions
like "voiced - unvoiced", "lips open -- lips closed", etc. So one could
look at a phoneme as a list of its values on each of these dimensions.

Yes, they do this. But Bruce has been trying to get you to see that this
is an unproductive way to look at phonemes. As he has been pointing out,
these features are continuum variables, not variables with binary values.
Linguists have categorical perceptions of them--lip opening has the state
"open" or "closed" for example. That's their perception, but it is not the
basis of the discussion we have been having.

So one could
look at a phoneme as a list of its values on each of these dimensions.
Putting the lists for two phonemes next to each other, one could see
how the phonemes "contrast" with one another.

This, again, is what Bruce and I have been trying to show cannot be true.
When the features are measured on a continuum, one finds that the same
set of values can represent either of two (or more) phonemes.

If the phonemes
are actually heard as different (contrasting) they must differ
(contrast) on at least one distintive feature (dimension).

(Take out "distinctive). Yes, but that dimension, in my view, relates
to the perception of the partner, not the acoustic signal.

The linguist
who uses this scheme can see how the phonemes differ (contrast) but
that doesn't mean that we percieve phonemes by actually making
the comparisons made by the linguist.

I don't think you will get much disagreement from speech researchers
(possibly as opposed to linguists).

Martin