🚀 go-pugleaf

[From Bill Powers (950217.1410 MST)]

Bruce Nevin (Fri 930217 13:24:53 EST)--

RE: applying the Test to language

Some questions, observations, etc. with the aim of clarifying the
proposal. Not intended as second-guessing or criticism -- just what
comes to mind as I read.

Procedure:

     1. We record what informants consistently tell us are repetitions
     of the same utterance. ...2. We note features of these instances
     that are different.  Because native speakers consistently judge
     these utterances to be repetitions, these differences are by
     definition incidental byproducts of control of a perception of
     different words.

If we consider perception to be a function of multiple variables
involved in utterances, "sameness" and "difference" are not predefined.
The Test involves hypothesizing functions of the observed variables and
proposing that the value of the function, not the individual variables,
is under control.

Suppose there are variables x and y defined at the physical level --
aspects of a sound spectrogram, time-domain measurements, etc.  Suppose
further that we find that repetitions of an utterance are perceived as
the "same", while we observe that these two underlying physical
variables differ over a series of instances. The problem then becomes
one of finding a function of the variables that is invariant with
respect to the changes that are observed -- so the function retains the
same value as the individual variables change in the observed way.

As an example of what I mean (not intended to be connected to speech
recognition), suppose we observe that variables x and y vary in the
following way:

  x     y

 10    99.5
 30    95.4
 50    86.6
 70    71.4
 90    43.6
100    00.0
...
-50   86.6
...
 30  -95.4

The person perceiving something derived from these variables claims to
perceive "the same thing" in each case. The problem is to find a
function of x and y that has the same value for each of these pairs of
variables.

In this case the function is p = sqrt(x^2 + y^2). All the above pairs of
variables yield a value of p of 100 when plugged into this function.
What the person claims to perceive as the same could be described as
"the distance of point at x,y from a point at 0,0." All these points lie
on a circle with a radius of 100.

A different set of points would lie on a circle with a different radius,
so it is possible for the perception either to remain the same or to
change. If we disturb the two points x and y in such a way that both the
initial and final pairs of values satisfy the equation p = sqrt(x^2 +
y^2), the person will experience no change in the perception, and will
not take any action to alter the points x and y. So we will see that
what appears to us as a difference in the environment is perceived by
the subject as the same. This will also be true if we hypothesize the
wrong function -- say, p = x + y. If we start with x = 50 and y = 86.6,
which would yield a value p = 50 + 86.6 = 136.6, we could change x and y
to 10 and 126.6, which has the same sum. We would expect the subject to
say "no error." But since the subject is perceiving sqrt(x^2 + y^2), the
actual value of p, which was initially 100 (sqrt(2500 + 7500) = 100),
will have changed to sqrt(100 + 16027) = 127. The subject will indicate
that a difference exists, or change x or y to correct it.

It isn't necessary to fit an analytical form like sqrt(x^2 + y^2) to the
observations. We could simply list the points that are said to be "the
same" in a table of x-y coordinates. This would provide a table from
which we could read off all the pairs of values that are perceived as
the same. We would then know that if we disturbed x and y together or
separately in a way that carried us from one pair of entries in the
table to another, the subject would say "no difference." Or if we
disturbed one variable and let the subject adjust the other to keep the
perception the same, we know that the other value would be chosen so we
were still in the same table of paired entries.

By introducing a third phsyical variable we could create a surface, a
two-dimensional table of triplets in three-dimensional space, and so
forth for as many dimensions as we could handle. Of course it is always
handier if we can write an equation to approximate the function, because
then we don't have to deal with large tables of entries, but it's not
necessary.

     3. We produce other pairs of different utterances by substituting
     one segment for another.  In place of some segment in one of the
     recorded utterances we substitute some perceptibly different
     segment from some other utterance.  We test these with native
     speakers as above to see if they are pronunciations of the same
     word.  (One or both may be nonsense syllables.)

Instead of just using pairs, it might be possible to use a whole series
of segments that differ systematically in one or more dimensions. Then
when a pronunciation is given, the subject can run up and down through
the list of varying pronunciations (with the sound being repeated each
time a new pronunciation is selected) until the subject is satisfied
that the original pronunciation has been restored. It would be
particularly useful if one aspect of the pronunciation could be changed,
while the correction is made by altering a different aspect of the
pronunciation (like altering x above and having the subject change y
until the perception is the same again).

     4. Result  The segments can be used as a representation for the
     contrasts between different utterances.

Each pair of segments so found represents a difference in a perception
due to a change in underlying physical variables. However, two
differences do not necessarily represent the same changes in the
underlying variables. That is, one utterance may be declared different
from another, on successive trials, for different reasons.

     Different analyses into segments are possible, such that each
     inter-utterance difference is in one-one correspondence with a
     segment, and vice versa.  The different analyses into segments may
     differ as to their practical usefulness for different purposes.
     Some are more efficient and elegant from a mathematical point of
     view, but unwieldy for a practical orthography, for example, a
     purpose that some other analysis into segments serves better.

In general, the inter-utterance differences will depend on more than one
dimension of difference in the underlying variables. If you have
variables a, b, c, and d, and establish a difference by varying c, you
might find that after changing b, you may not get a difference after the
same change in c, or you might even reverse the difference due to a
change in c.

If you give the subject control over one or more components, you can
make a map of all the quadruples of values of a, b, c, and d that are
perceived as "the same" utterance. As you vary the components you will
get a list of "contrast" judgments and "no contrast" judgments; both are
important. The "contrast" judgemnts tell you that you have moved from
one four-dimensional region into another; the "no contrast" judgments
tell you you have moved to a different place in the same region.

     The pair test and substitution tests demonstrate that the speaker
     and hearer control perceptions of identifiable points of difference
     between words.

I think we are basically talking about the same thing. I've tried to
express the procedure in terms I'm used to, with perceptions being
functions of multiple input variables that take on different values.

     The fact that there is no unique analysis into segments, and that
     alternative analyses are equivalent as representations of these
     identifiable points of difference between words (and as
     representations of whole words and utterances, of course), suggests
     that speakers control perceptions of the contrasts between words.

I might quibble here in that perceptions are functions of the _values_
of the underlying variables, with differences or contrasts merely
expressing the change from one perception to another. By casting this as
a multidimensional problem, we can see that there will be strong
interactions among the underlying variables, so a contrast produced by
varying one component will depend on the values of the other components.
However, this is a quibble; the method of contrasts will probably
provide a map of the perceptual space involved, particularly if you use
the "method of adjustment" to establish sameness as well as difference.

     The suggestion that speakers control perceptions of the contrasts
     between words is supported by the diversity of means that may be
     used to effect a given point of difference between words, and the
     diversity of perceived differences that a given choice of means may
     effect.

I take it that this is not something that linguists are in agreement
upon. Is this the main motivation for the study, as far as linguists are
concerned?

     It has been a fundamental tenet of Generative linguistics that
     there is a universal set of features that may be used as means of
     contrasting utterances in any given language.  Not all problems
     encountered in this project have been resolved.  Nonetheless, a
     more or less standard set of distinctive features is assumed in the
     literature.  It is understood that there may be a diversity of
     phonetic means for effecting a given phonological feature,
     differing between languages and differing between phonological
     contexts (as above) within a given language.

One remaining problem is that in measuring "features" you are already
running the raw physical data through a set of perceptual functions to
yield the states of the features. If there are problems in applying the
method of contrasts, it may be that the underlying features need to be
derived in some other way. The sound spectrograph may not be the ideal
way of getting variables like the lower-level variables involved in
human speech recognition. But I don't mean to borrow trouble; it will be
interesting to see how far you can get with the approach as described.

     Means are needed for disturbing features of speech while providing
     means for native speakers to resist those disturbances.

Roger. I'll send you some material on the Sound Blaster Developer's Kit,
which describes what can be done with this card. I had no difficulty in
becoming a Creative Labs developer, the kit costing about $110 and the
card about the same (for a developer). I just wrote a letter of
application and filled out a form that they sent.

Creative Labs
1901 McCarthy Boulevard
Milpitas, CA 95035
Attn: developer relations

Are you going to be working with Avery Andrews on this? I guess it's
time to start talking about the equipment you will need and figuring out
how to cover the costs. We'll consult the CSG officers and look around
for other contributers, too.
-----------------------------------------------
I can see you're a bit pissed off with me about the "social phenomena"
discussion. Can't say as I blame you. I'm trying to get across a point
of view about explaining things, and I have never been able to do it
coherently. It seems to me that in speaking about "social
responsibility, reliability, interdependence, dependability, and so on"
we just aren't talking about universals in the same way we are when we
talk about controlling perceptions. Those social things seem much too
general and indefinite. They seem more like ways of classifying
behaviors than theoretical concepts, sort of like classifying aptitudes
and preferences. I guess that "society" has never seemed as real to me
as the people in it, and I've seen so many different ones that it's hard
for me to grasp what's the same about them as societies. I'm just not a
generalizer; maybe I'm a sociopath, too.

Anyway, don't take my obtuseness too hard. At least you know that
whatever generalizations you come up with concerning social phenomena,
I'm not likely to propose a rival theory.
-------------------------------------------------
     I suggested farther on that secretion of endorphins (endogenous
     morphines, I am told among the most powerful opiates known) could
     well be an innate mechanism involved here.  I understand (can't
     cite sources) that endorphins are released e.g. when one animal
     grooms another and when animals play, two obvious ways of affirming
     reliable relationships.

What releases the endorphins? Maybe it's the same system that initiates
grooming, etc..  I'm not much impressed by chemical explanations of
human experience and behavior: they always seem to assume that once
you've found a chemical that's involved somehow, you've found a cause.
Why ask where it came from, and why?

What's obvious about grooming being a way of "affirming reliable
relationships?" That's much too high-falutin' a way of talking about
what monkeys do. Something affirms reliable relationships only if you
perceive it as doing so, and know what "affirming," "reliable," and
"relationship" mean. And agree that that's what's happening.

     In play, animals do things that could be perceived as attack and
     fighting, but refrain from injuring; the effect is to communicate
     "I could attack you, I could be your enemy, but I am not".

More of the same.

     The use of snapping and growling varies a lot from one culture to
     another.

And also within "one" culture.

     Ruth Benedict (anthropologist) introduced the term synergy,
     borrowed from medicine, as a descriptive term for this.  At one
     extreme, in a high-synergy culture there are no words and evidently
     no contrastive concept of selfishness vs. altruism.  In such a
     culture, it is self-evident that one cannot benefit another without
     corollary benefit to oneself, and that one cannot injure another
     without corollary injury to oneself.

Nothing is self-evident until you learn it. Even if something is true,
you do not know it is true until (and unless) you specifically learn it.
Synergy is a word about an imaginary situation; there are no "high-
energy" and "low-energy" cultures -- people in a culture are just too
different from each other to be classified so neatly. Ruth Benedict
evidently likes simplifying generalizations. I don't. I really don't
think there is anything you can say truthfully about "people" at this
level of abstraction.

Sorry. This style of explanation is not my cup of tea. I guess I should
just shut up about it -- I'm sure all I'm succeeding in doing is giving
you reason to be irritated at me.
------------------------------------
     After a couple of turns of the conversation, he told me something
     that startled me.  He said that his first reaction when he
     encounters another man is, "can I kill him?"  Can I overpower him.
     Fight or flight, as he later put it.  He claimed that, based on
     conversations with other men, he thought this was universal.

Well, he was wrong, wasn't he?

     If he knew this, he might say that my choices in the matter,
     avoiding conflict, reflect my recognition that I could not kill
     him.

This is known as sophistry. Sure, people can say anything they please.
Do you believe that everything people say to you has to be taken
seriously? I think that if you wanted to kill someone you would know it.
The greatest tool of the charlatan is to claim that he knows something
about you that you don't know. Gigabucks of profit in that.

At least this fellow told you something useful: now you know what he
thinks when he meets you. I'd avoid him; he might mean it.
-----------------------------------------------------------------------
Best to all,
Bill P.