NetNews Usenet Archive 1993 #3/NN_1993_3.iso/spool/bit/listserv/csgl/2320

Text File | 1993-01-28 | 12.3 KB | 261 lines

Comments: Gated by [e-mail]
Path: sparky!uunet!uvaarpa!darwin.sura.net!paladin.american.edu!auvm!CCB.BBN.COM!BNEVIN
Return-Path: <@VMD.CSO.UIUC.EDU:[e-mail]>
Message-ID: <[e-mail]>
Newsgroups: bit.listserv.csg-l
Date:         Thu, 28 Jan 1993 11:31:01 EST
Sender:       "Control Systems Group Network (CSGnet)" <[e-mail]>
From:         "Bruce E. Nevin" <[e-mail]>
Subject:      theories of language
Lines: 249

[From: Bruce Nevin (Thu 930128 08:29:27)]

(Bill Powers (930127.1430) ) --

>Interesting. Here we are again. Operator grammar explains
>language. Generative grammar explains language. Pick one to
>believe, and the other is wrong. Obviously they can't both be
>right. What basis can there be for anyone to believe either side?

Neither "explains language" in a way that is satisfactory from a
PCT perspective.  Either could be used as a starting place for
outlining a PCT explanation.  My argument is not so much that
Generative theory is wrong, but that it is *much* more difficult
to go anywhere using it as a starting place.  (I would argue that
various flavors of Generative theory are wrong about certain
things, but those things are by and large irrelevant or secondary
to our concerns here.)

Let's go back to your (930126.0830) generalizations:

> One of the logical errors that's easy to make in learning control
> theory is to suppose that because a particular state of the
> environment is observed to be involved in control of a higher-
> level perception, ONLY that state of the environment will result
> when the perception is controlled.

> I've seen this in linguistics. In a top-down model, some global
> feature of a sentence is specified. This feature is then
> exemplified by some element of a specific sentence at a lower
> level. But why that sentence, and not a totally different one
> that is also an example of the higher form? In fact the detailed
> sentences used as examples vary all over the place, so there is
> clearly no constraint on which sentence is to be used as an
> example. This is a major problem for top-down models (at least as
> far as implementing them is concerned).

> A control-theoretic model of language production works the other
> way: it selects sentences until one is found that can be
> PERCEIVED as having a certain form matching a specified form.
> None of the degrees of freedom of the sentence matter except the
> combination that results in satisfying the reference-form; in all
> other respects the sentence is free to vary. Any old words that
> can be perceived as an NP will satisfy the requirement that an NP
> appear in the final result in that position. But you can't go the
> other way: you can't write an algorithm that will start with the
> specification that a noun phrase be uttered, and come up with a
> specific utterable noun phrase. The specification "NP" doesn't
> care which noun phrase is found; therefore it can't specify ANY
> noun phrase.

> You linguists out there have heard me harp on this before. So far
> you haven't dealt with the basic problem. It's part of the same
> problem that Rick is talking about.

The segue from your reply to Rick's (930125.1500) on output and
perception "bushes" is actually a non-sequitur.  A
phrase-structure tree with NPs, VPs, etc. is a structural
description that applies to an indefinitely (but not infinitely)
large set of sentences.  No one has ever claimed that one gets
from a phrase-structure tree to a particular sentence in that
set, top-down.  That is not what the term "generative" means.
(Basically, it just means "explicit," aside from its role as a
trademark.)  Indeed, that is one of the conceptual difficulties
of "lexical insertion" in classical transformational-generative
grammar.  How do the words come in, that is, with what
motivation?

The bottom-up control-theoretic approach that you counterpose to
this is I think what Avery has been proposing.  (Correct me if
I'm wrong, Avery, or elaborate if you will.)  The words come up
anyhow, in association with non-word perceptions.  There are
certain dependencies among the words that are correlated with
dependencies among the perceptions.  (These dependencies among
the perceptions do not themselves all necessarily have to be
perceived for such correlation to come about, I believe.)  But
without further constraint they might come out in a way that was
not English (or the given language).  This further constraint
comes from perceptual control with reference to the syntactic
structures.

(Penni proposed that you don't even need the syntactic
structures.  This is because some aspects of structure,
particularly discourse structure, follow from the correlation
with dependencies among non-word perceptions.  You can seem to
get away with it for simple texts like hers if you build in
English word order without noticing that you have.  A simplified
sublanguage like this could be of considerable practical value in
computer applications prohibiting dialog or limiting the
complexity of what the user could input.)

With a Phrase-Structure basis, the programs to generate all and
only the permissible structures needed to provide these reference
perceptions are quite complicated.  They require control of
perceptions that are quite abstract.  It has been argued
convincingly (translating into PCT idiom) that infants cannot
possibly develop perceptual control of such complicated and
abstract perceptions on the basis of generalization and
abstraction from their observations of language in use around
them, and that therefore the hairy parts of it must be hard-wired
in the human genome.  This is known as the argument from paucity
of data, or from paucity of stimulus.

With a word-dependency/reduction basis (operator grammar), there
are no abstract structures.  Start with the word dependencies, as
above.  The programs for reducing the output form of
low-information words are simple.  The programs for doing so in
conformity with current convention are a bit more complex in the
places at which they are arbitrary, as are the programs for
strategies to avoid various pitfalls, but no more complex or
difficult to acquire than those for many sorts of nonverbal
control.  The ill-defined and variable set of sentences results
as a byproduct of these several kinds of perceptual control,
rather than by control with respect to a well-defined set of
abstract sentence structures.

A PCT explanation of language can be undertaken with either
approach.

>                                In Operator grammar you need to
>know what words mean in order to distinguish operators from
>arguments

No, you don't.

>Suppose I were to give proponents of either side a sentence like
>"Word1 Word222 Word17 Word9237 Word1403 ...".

No one can develop a grammar on the basis of one sentence.  This
is in no way different from being handed something in a language
you don't know, say, this sentence from an article on the front
page of a Greek newspaper on my desk now:

        I apoxorisantes i ekdhyoxthendes ipurghi ke poli vuleftes
        thetun evtheos ke aprokaluptos to erotema: pera apo
        linovouleftikus tipus, pu dhinun iperoxi dhio psifon, me
        pya praghmatiki pliopsifia kiverna ti xora o prothipurgos
        k. K. Mitsotakis.

It is only by finding commonalities across many utterances,
starting of course with short ones, that one distinguishes word
classes, and that is on the basis of what can co-occur with what.
One would find many instances words ending in -es in proximity,
often with the word i or tis before them, and others ending in
-os with o before, and so on.  One can identify affixes and words
and their gross combinatorial possibilities (word classes)
without any translation.

With a translation, one can use one's control of one's own
language (and its metalanguage) as a basis for controlling the
new one.  Naturally, this results often in conflict.  Infants are
not troubled with this sort of conflict, and that, together with
the fact that they work at it pretty near full time for a couple
of years, integrated with learning all sorts of perceptual
control, accounts for the difference between child language
acquisition, and Genie, and adult second-language acquisition.

The partial order of word dependencies, operators over arguments,
before imposition of a linear order of words, derives from
control of meanings (correlation with dependencies among non-word
perceptions).  But one's growing control of what words can be
placed where in linearized order is based in large measure on
language-specific convention (for the linearization of the
operataor word with respect to its argument words, plus a few
analogic extensions).  The remainder falls out as a byproduct of
reductions.

>we expect the model to work correctly no matter what specific
>meanings we give to these variables.

>A model of language constructed in the same spirit as the PCT
>model would not need a lexicon or empirical data on the way
>specific words are used.

The meanings are non-word perceptions with which words and some
phrases (speaking loosely) are associated.  They are inside the
model insofar as control involving the non-word perceptions is
modelled.  The "lexicon" comprises the word-perceptions, the
association of words with non-word perceptions, the control of
reductions in word shapes (morphophonemics).  These are all
perceptions whose control is modelled in the model.  If we model
language learning, then empirical data are outside the model and
inside the model.  They are

>                                                needed to
>arrive at such a model, but once the model was constructed it
>would no longer be cast in terms of specific observations.
>Instead, there would be underlying principles that apply to any
>way of hooking symbols to experiences, whatever the symbols and
>whatever the experiences. In fact I doubt that this underlying
>model would be a linguistic model at all: it would simply be one
>application of a single model of perceptual control, the same
>model that explains all behavior.

I still don't understand why you believe we disagree about this.
Try again?

>It's this concept of an underlying model that seems to me to be
>missing from both the Operator and the Generative approaches. If
>the apparent laws of language depend on specific word meanings,
>they are not laws of language but only happenstance. Any other
>laws would be just as likely, for all we can explain why they
>exist.

This concept of an underlying model can be added to either approach.
The results in operator grammar do not depend on specific word
meanings; rather, an account of meanings of words (and of
sentences, and of texts) is a result in operator grammar, an
account that fits well with PCT.  Some aspects of language
structure are due to correlation of word-perceptions with
non-word perceptions.  Is this what you are rejecting here?

In Generative Grammar, much has been made of the autonomy of
linguistic form from meaning, though Chomsky has been backing
away from that recently.  Still, it is not clear to me how you
see that the "apparent laws of language" according to Generative
theory depend upon specific word meanings.

(Avery Andrews 930128.1120) --

>There's something about the Harrissian approach that I just don't
>get, & until I either get or it figure out what is wrong with it
>I have no intention of saying much of anything against it.

>In general, I think that current linguistic theories should be regarded
>as just being resources that are out there, and that from a PCT
>perspective, one should just take whatever insights, if any, that
>they might seem to offer.

Harris has been ignored for 40 years for (academic-)political
reasons.  That's a shame.  It makes his work effectively
unavailable as a resource and makes his insights invisible.  If
you or anyone shows me something that is wrong with it, that will
be a good outcome.  I'll just keep putting it out there and
trying to make it as vulnerable as possible.  If you can outline
what it is that you don't get, I might be able to help you expose
it better.

>                          And, at the moment, I see Motor Control
>as a *much* higher priority target.

I agree.  I don't see practical modelling of language being
feasible unless control up to category level is modelled or
simulated.  I doubt that practical modelling of control at
category level or higher is feasible without modelling language
or covertly smuggling language-borne understandings into the
model.

        Bruce
        [e-mail]