NetNews Usenet Archive 1992 #26/NN_1992_26.iso/spool/bit/listserv/csgl/1402

Internet Message Format | 1992-11-08 | 28.2 KB

Path: sparky!uunet!stanford.edu!bu.edu!wupost!darwin.sura.net!paladin.american.edu!auvm!CCB.BBN.COM!BNEVIN
From: [e-mail] (Bruce E. Nevin)
Newsgroups: bit.listserv.csg-l
Subject: language issues
Message-ID: <[e-mail]>
Date: 6 Nov 92 19:11:10 GMT
Sender: "Control Systems Group Network (CSGnet)" <[e-mail]>
Lines: 576
Comments: Gated by [e-mail]

[From: Bruce Nevin (Mon 92112 07:38:16 through Fri 92116 14:05:06)]

Awfully hard getting any consecutive time to respond to all the
interesting stuff going by.

(Avery Andrews 921026.0928) --

>I'd agree that this little story of mine would stand or fall on the
>basis of careful examination of this kind of case.  But if this story
>falls, we're probably left with something like GB as the best bet for
>a theory of grammar!!.

Are the choices really so few and so constrained?
This seems based upon an argument of the following form:

    Either A or B.
    If A, then something like GB
    If not (this little story of mine) then not B

Here, A is something like the claim that "learners don't have access to
negative evidence," i.e. they are not provided with starred examples,
and B is the claim that they are.

For non-linguists: in Generativist writings one puts an asterisk before
examples that "you can't say" (at least according to the author).
Hence, these are "starred" examples.  (I used to call them "asterisky.")
For example, the linguist might consider sets of roughly equivalent
sentences like this:

     1a     John gave the Ramadan bake sale money to the fund.
      b     John gave to the fund the Ramadan bake sale money.
      c     John gave the fund the Ramadan bake sale money.

On the basis of this and other sets of sentences, the linguist might
postulate a rule that says you can move a "dative" phrase--"to x" after
a verb with some kind of "give" meaning--to a position before the verb,
and delete the preposition "to".

This rule predicts that the following sentences are also equivalent to
each other:

     2a     John donated the Ramadan bake sale money to the fund.
      b     John donated to the fund the Ramadan bake sale money.
      c    *John donated the fund the Ramadan bake sale money.

However, for many English speakers (including Avery), (2c) is not
acceptable, hence the asterisk or star.  Others, focussing on the
meaning of (2c), may overlook the syntactic or stylistic anomaly; Bill
is evidently among these, but it is possible that he represents a third
set of others for whom (2c) is not anomalous at all.

Returning to the issue at hand:  the claim that language learners do not
have access to negative evidence seems to amount to a claim that they do
not experience error when they produce syntactically anomalous
utterances.  The claim is usually that overt correction ("don't say X,
say Y") is such a miniscule part of the language learner's experience,
and covers so few of the possible errors, that it alone cannot account
for the incredible amount of learning that takes place in such a short
time (hence much of the structure of language cannot be learned at all
and instead must be hard wired in the genome, etc, etc).

Several assumptions underly this claim.

Doesn't this assume that overt correction by other speakers is
the only source of negative evidence?  It should be obvious to students
of PCT that there are many occasions for experiencing error aside from
attempts of others at social control.  For starters, consider the
experience of not being understood and having to try again, or the
experience of having the other paraphrase what one has said, either in
confirmation or in replaying it to someone who did not understand
clearly, both of which are *very* common experiences for children
learning a language.

Generativists assume that the learner is faced with a range of
alternative grammars, and that learning is a process of eliminating
those that don't work.  On this view of how learning works,
counterexamples have an enormously important and pervasive role.  GB
(Government-[and-]Binding Theory, now being supplanted by Chomsky's
latest "minimalist" proposals) carries this to an extreme.  UG
(Universal Grammar), hard wired into the genome, provides for all
possible languages.  Learning one particular language involves setting
parameters in UG so that one possibility is admitted for that parameter
and the others are precluded.  An example might be the choice of SOV
(subject object verb) word order, as opposed to the alternatives.
Having made that choice, certain other choices ripple out as entailed
consequences (order of modifiers relative to modified words, etc.).
Much ink has been spilled in a search for the minimal parameters whose
settings cannot be predicted from the settings of other parameters, or
which once set can be the basis for predicting a larger number of
others, etc.  (In this research climate, language differences and
variation are not always so closely investigated as they might be, but
that can be only a secondary side comment here.)

Generativists also commonly perceive that language structure is so very
complex and arcane as to defy a child learning it at all absent
biological inheritance of much of the complexity in UG.  It is true that
Generativist descriptions of language are very complex and arcane.  It
is also true that not all ways of describing language are so.

None of these assumptions is very convincing to me, but they do keep a
number of people employed as linguists (and a number of others not
employed as linguists).  It also puts a great many properties of
language in genetically determined UG and therefore outside the realm of
perceptual control.

********************

Returning to the examples, sentence (2c) is somewhat peculiar for me,
but not completely unacceptable, so I mark it with a question mark
rather than an asterisk:

     2c    ?John donated the fund the Ramadan bake sale money.

Contrast (2c) with an example that unequivocally merits an asterisk:

     3. *John fund the donated.

In operator grammar, the "to" may be zeroed with a verb after which it
is so strongly expected as to be redundant.  (In some UK dialects, it
need not be immediately after the verb, e.g. "John gave it me."
Similarly in Danish.)  Thus, the reason given in Operator Grammar that
(2c) is peculiar is that speakers of English do not have a sufficiently
strong expectation of the preposition "to" occurring after "donate."

One might suppose this is because "donate to" simply is not used as
frequently as "give to."  It is true that the dative reduction is more or
less peculiar for many other verbs in the "give" set.  Call them the
"donate" subset, including:

? John administered [to] the fund the money
? John allocated [to] the fund the money
? John apportioned [to] the fund the money
? John communicated [to] the fund the money
? John consigned [to] the fund the money
? John contributed [to] the fund the money
? John conveyed [to] the fund the money
? John dealt [to] the fund the money
? John delivered [to] the fund the money
? John dispensed [to] the fund the money
? John distributed [to] the fund the money
? John entrusted [to] the fund the money
? John purveyed [to] the fund the money
? John relinquished [to] the fund the money
? John rendered [to] the fund the money
? John surrendered [to] the fund the money
? John transferred [to] the fund the money
? John transmitted [to] the fund the money
? John vouchsafed [to] the fund the money

But a considerable number of other verbs in the "give" set are
acceptable with the transposed argument order and zeroed "to".  Call
them the "give" subset:

John accorded [to] the fund the money
John allotted [to] the fund the money
John assigned [to] the fund the money
John awarded [to] the fund the money
John conceded [to] the fund the money
John furnished [to] the fund the money
John granted [to] the fund the money
John handed [to] the fund the money
John left [to] the fund the money
John lent [to] the fund the money
John offered [to] the fund the money
John paid [to] the fund the money
John presented [to] the fund the money
John proffered [to] the fund the money
John provided [to] the fund the money
John sent [to] the fund the money
John supplied [to] the fund the money
John yielded [to] the fund the money

I don't think a "frequency of occurrence" argument holds up.

Now, according to Bill's intuitions about how language works, we use
words because they are associated with perceptions that we are trying to
get our audience to attend to (in the environment or in imagination).
Examining these two lists, I don't see any obvious differences in some
sort of "to" perception associated with one more strongly than the
other.

Instead, the difference appears to be a pretty arbitrary, learned,
socially shared perception that "to" is redundant after verbs in the
"give" list, but not after those of the "donate" list.  Maybe someone
else can pull out some generalization that I'm missing.

There is another factor, one that motivated my putting the modifier
"Ramadan bake sale" in the examples of (1) and (2), and that is that by
minimizing the distance between an operator and its arguments (distance
in terms of intervening words that could also be candidates as
arguments), we reduce the load on short-term memory for perception of
operator-argument dependencies.  Thus, when we have alternative
orderings for the arguments of an operator we prefer to put the shortest
one first (counting all its modifiers):

    4a John handed the money over to the fund.
     b ?John handed over to the fund the money.

Sentence (4a) is definitely better than the transposed version (4b).
However, watch what happens when you lengthen "the money" with some
modifiers:

    4a John handed the Ramadan bake sale money that Kathy had collected
       from the first half of her list of groups over to the fund.
     b John handed over to the fund the Ramadan bake sale money that
       Kathy had collected from the first half of her list of groups

Now consider the original two sets of sentences without the modifier:

     1a'  John gave the money to the fund.
      b' ?John gave to the fund the money.
      c'  John gave the fund the money.

     2a'   John donated the money to the fund.
      b'  ?John donated to the fund the money.
      c' ??John donated the fund the money.

Notice that (1b') is a bit awkward but (1b) is more acceptable:

     1b' ?John gave to the fund the money.
     1b      John gave to the fund the Ramadan bake sale money.

This is seen in the "donate" sentences as well:

     2b' ?John donated to the fund the money.
     2b   John donated to the fund the Ramadan bake sale money.

The fronting of the second argument of "give" (with its argument-
indicator "to") is motivated by its length with its modifiers in (2b),
but there is no such motivation in (2b').

Control for minimizing the distance between operators and their
arguments is presumably universal, but to say it was a property of
Universal Grammar would be fatuous.  It is surely a universal property
of control systems that are controlling for sequences that may be
interrupted by other controlled sequences.  So likewise, I believe, many
other claims for UG are really claims for the universality of control.

(Bill Powers (921024.0830) ) --

> Avery Andrews (921024.1519) --

> >So the non-occurrence of the second (a piece of `negative
> >evidence') becomes accessible to the learning system.

> Control theory handles the non-occurrance in terms of a reference
> signal that demands the occurrance. When there is a reference signal
> spcifying that some perception occur, then as soon as the reference
> signal is set there is an error, which persists until the perception
> occurs. In this case it is not occurrance of the sentence that
> matters, but of the meaning.

Maybe you've sorted it out by now.  You and Avery were operating at
right angles to each other throughout much of this exchange.  In this
instance, "non-occurrence" meant for Avery a starred example--something
marked as not acceptable or not occurring in acceptable usage, or
perhaps (by the reification that is normal in Generativist literature)
not occurring in the grammar.  At issue above is the question who if
anyone other than the linguist has a reference signal for such a thing
occurring; and it is not the meaning of the sentence that matters, but
whether or not you can express the given meaning in that particular way.

(Avery Andrews (921024.1519) ) --

>A further prediction is that the kind of optionality above will be
>an inherently unstable feature of languages:  if two such forms are
>used with no discernable difference in meaning, the language-
>acquisition systems of the speakers will be constantly reorganizing
>without being able to find an error-free configuration, &
>presumably at some point one of the forms will other drop out, or
>they will acquire subtly different meanings

Something like this was advanced by Kurylowicz, re the classic
observation that "doublets" (specifically an inherited but "irregular"
form alongside the form used by an innovating group, not necessarily a
younger generation, that used analogy to create a more "regular" form)
tend to become differentiated semantically.  I think this is part of a
more general process whereby distinctions tend to be exploited as
differences that make a difference, or else are no longer maintained.
But this is a *social* process (i.e. ongoing negotiation of agreements
about what a given distinction either constitutes or means).  Your
formulation here does not differentiate between this social process and
the processes of individual language learners, hence Bill's inference:

>The implication is that the mature speaker will come to express the
>same meanings in the same words all of the time, never paraphrasing or
>varying the wording. I'm not sure you would want to maintain that.

What prevents rapid closure, smoothing out all the bumps and ripples in
the process of learning a language, is precisely the fact that it is not
a process that any one person can carry out in isolation, but rather a
social negotiation of agreements, and the prior commitments and
motivations of the participants are not necessarily commensurate with
each other.  Pronouncing "bird" as "boyd" constitutes a token of
membership in a community that talks that way, and presenting oneself to
others by using an ensemble of such tokens means different things to
different people, and on different occasions.  A person who can and does
pronounce "bird" both ways may in fact be motivated NOT to smooth out
the difference.  (Use one in Brooklyn, use the other in my job as a
newscaster.)

As you say,

>if two such forms are
>used with no discernable difference in meaning

But this is to say that they are used without ever experiencing error,
no matter which is used.  You suggest that the availability of
alternative ways A and B of expressing meaning M, where A and B are not
distinguished from one another (no A-circumstance where it is an error
to use B, and vice versa), is itself an occasion for error.  But this
is a sort of meta-error, of a different logical type from the the first
that you proposed, as I understand it:

1. A person hears utterance x
2. The person's "comprehension system" comes up with meaning m.
3. The person's "production system" generates in imagination a set of
   utterances U = {w, y, z}, where each member of U expresses meaning m.
4. The absence of utterance x from U is occasion of an error signal.

Apparently, the person guesses at the meaning m.  There must have been
some basis for this guess--word choice, context, part of x is like part
of w, part is like y, and so on.  It seems to me that what one typically
does with this sort of error is to perceive the other person's utterance
x as an error of production or reception or both, perhaps as a repair or
a change of syntactic horses in mid-sentence, etc.  One's conversational
response (or one's strategy as a listener or reader) includes heightened
attention to those elements (words, word dependencies) about which one
is less certain.  One attends also to some alternative meaning m', and
its alternative utterance set {p, q, r} also generated in imagination in
parallel.  One of these may suddenly pop up to most favored status.
Socalled "garden path sentences" provide simple examples of this
process: "The horse raced past the barn fell down".

In the meta-error case, the person wishes to express meaning m and finds
that there are alternative ways {w, y, z} for expressing it.  If there
is some motivation for choosing one rather than another, some difference
that makes a difference to the person, then the choice is obvious and
there is no error.  Commonly, it seems to me, the unthinking preference
is a repetition of words or of a syntactic construction recently used in
the same discourse (conversation, speech, text, etc.).  A more polished
speaker may make a different choice with deliberate artifice, perhaps
for emphasis or simply for variety.  Again, so long as there is a basis
for choice, I don't see any occasion for error.

Error may occur for example when choice w vs. choice y constitutes a
token of social membership (like the different pronunciations of
"tomato," "neither," and "economics").  This is not a difference of
meaning in the sense that Bill intends, I think.  (That's why I say it
"constitutes" instead of "means" social membership.  It's performative,
in the same sense that "I hereby declare this a disaster area", spoken
by a duly empowered official, constitutes a change to the status that
the phrase "disaster area" by itself means.)

You (Avery) had specifically in mind examples like (2a') and (2c'):

      2a'  John donated the money to the fund.
      2c' ?John donated the fund the money.

You propose that just because the learner's "production system" at time
t produces both of these indistinguishably to express a given meaning,
then every time the learner hears (2a') an error signal results.  The
language learner's response, you proposed, is to eliminate that error by
reorganizing, specifically, by altering the production system so that it
no longer produces (2c').

I think that a one-step elimination, analogous to marking (2c')  with an
asterisk (or question mark), will not do.  Such a mechanism must surely
function to eliminate production of (1a') every time the learner hears
(1c') and to eliminate production of (1c') every time the learner hears
(1a'):

      1a'  John gave the money to the fund.
      1c'  John gave the fund the money.

Instead, perhaps there is an incremental change in the expectability of
one form or another.  Since there may be three or four or more
alternative forms in some cases, the best move it would seem to me is in
the opposite direction: to *raise* the expectability of the form that
actually occurs, rather than lowering that of the ones that do not
occur.  But no error signal is needed for this, and no reorganization.

In fact, I see no obvious reason to suppose that a construction such as
(2c') could not hang around for the life of the language learner--one of
many constructions predicted on analogy to forms that the learner has
heard ("predicted by the rules of the grammar" in the customary
hypostasis), but not itself actually encountered.  I think that it would
be untenable to claim otherwise in the face of language novelty,
language change, and our capacity for understanding the meaning of
utterances across dialect variation, non-native usage, speaker error,
difficulties of hearing, and so on.

This interpretation accords with the fact that the acceptability of
sentences is not a binary, yes/no property, but rather a graded
property, and probably complexly so (that is, on more than one graded
parameter in parallel, e.g. differentiated by subject-matter domain for
starters).

So I think there are other ways of showing that language acquisition,
like all control processes, is guided by perceptual error ("negative
evidence"), and that the alternatives are not limited to your proposal
and GB.

-=+=-=+=-=+=-=+=-=+=-=+=-=+=-=+=-=+=-=+=-=+=-=+=-=+=-=+=-=+=-=+=-=+=-=+=-

(Bill Powers (921025.1900) ) --

>                              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.

This diagram makes the point about arriving at outputs by continuous
control of perception rather than by top-down realization of some sort
of archetype.  The actual flow of control is probably quite different.

Penni asks (penni sibun 921026):

>where are the symbol structures coming from?  why are any necessary?
>what does meaning have to do with anything?  why can't speech actions
>be just like other actions, which presumably aren't mediated by
>``meaning''?

I would suggest that all perceptions "have" meaning, that is, any given
perception is linked in associative memory to other perceptions.  I'm
not sure precisely what you (penni) mean by the phrase "mediated by
meaning".  Let me lay out a bit of what I think is going on.

Even before we consider meanings, language perceptions involve most of
the proposed levels of the perceptual hierarchy, if not all.  Phonology
alone runs from intensity level up to at least the event level, possibly
the program and even principle levels if some ways of doing phonology
survive the eventual PCT shakeout.  Morphemes and words are event
perceptions in Bill's usual estimate of things.  (Bill: Is this
problematic if syllables or semisyllables turn out to be events?  If so,
word events would be constituted of syllable events.)

Nonverbal perceptions obviously involve all the proposed levels of the
hierarchy, independently and in parallel to these language perceptions.

Event-level word perceptions are linked with nonverbal perceptions.
Bill has suggested that the input of a category perception ECS can be
satisfied by various nonverbal perceptions (glimpse wagging tail,
barking sound ==> dog) or by a word ("dog") or both.  I have suggested
that the link is in associative memory, and that it is not a simple 1-1
correlation of words and categories.

    Nonverbal Perceptions           Word Events

    P1 P1' P1"              <---->      W1
       P1' P16 P17          <---->      W2
    P2 P3 P4 P5
    P6 P6' P6"              <---->      W6
    P7 P8 P9 P10 P11 P12
    P13 P13' P13"           <---->      W13
    P14 P15 . . .

Here, I represent word W1 as evoking perceptions P1, P1', and P1", word
W2 evoking perceptions P1', P16, and P17, and so on.  This sort of
partial overlapping and intersecting is typical of word meanings.
("Dog" links through one sort of perception back out to the word "hound"
and through another sort of (sexist) perception back out to "broad."
Each of these words evokes other perceptions, which in turn have yet
other words associated with them.)  Conversely, perception P1' evokes
both word W1 and word W2.

Now suppose I want to communicate perception P1' to someone.  I can pick
either word W1 or word W2 (in our simplified universe here).  My choice
might be determined in part by extraneous perceptions associated with
one word but not the other.  As Penni and Avery have suggested, my
choice is constrained by considerations that apply to the words qua
words with respect to other words already spoken or planned to be
spoken.

Among these considerations are the operator status of the word, the
other words it requires as arguments.  Bill argues that these factors
are determined by the perceptions (meanings) associated with the words:
You can't have a "jump" perception without a perception of a thing
jumping.  Against this is the proposition that there are only certain
kinds of perceptions that you can talk about, namely, those provided for
in your language and culture.  (You can talk about more inchoate
perceptions, but only with difficulty, and with far less assurance of
being understood.)

Also among these considerations are the reductions available when the
word enters on its arguments.  This is determined largely by convention.
Some redundancy allowing for reduction is determined by universal
criteria such as word repetition, but as we saw earlier with respect to
the "give" and "donate" lists of verbs some of it is learned convention,
differing from one language to another, with a determinable history, and
with many regular correspondences among the different languages spoken
by people whose ancestors once spoke the same language.  One learns
these idiosyncrasies and controls one's perceptions of one's behavioral
outputs for conformity to them, for the sake of constituting one's
membership in one social group or another.

>                              desired
>                              meaning (nonverbal perception)
>                                |
>           perceived meaning-->COMP-------error --->
>                |                                   |
>                |                                   |
>          input function                       output function
>            |  |  |  |                           | | |  |
>            sentences                         sentence-variations
>               |                                     |
>                <---Imagination or real words <-----

In this loop, you have to understand that the meanings are not on a
higher level of the perceptual control hierarchy, they are in parallel
nonverbal parts of the hierarchy.  The labels "input function" and
"output function" conceal a great deal of complexity controlling for
words as event perceptions, operator-argument dependencies, reductions,
and "style" perceptions of many sorts (ranging from "that's a cliche" to
"what kind of person talks this way").  And as Martin emphasizes, the
"desired meaning" component probably equates to one's perception of
the audience "getting it," in imagination or in the environment.

I'm sorry, I can't seem to do any better than that just now, and I see
by my date stamp that it has taken me all week to get even this far.
Maybe I can do better next week.

-=+=-=+=-=+=-=+=-=+=-=+=-=+=-=+=-=+=-=+=-=+=-=+=-=+=-=+=-=+=-=+=-=+=-=+=-

This is what I started with, but I'm putting it last as being of least
interest to non-linguists.

(Avery Andrews 920211.1049) --

The terms S, NP, `noun phrase', and `clause', and the familiar tree
structures (and rewrite rules, labelled bracketings, etc.) are elements
of certain systems for describing language, based upon phrase-structure
grammar.  Insofar as they are not elements of other systems for
describing language, and absent any demonstration that the
phrase-structure systems are "correct" and the other systems are
"incorrect" with respect to the status of these elements, it seems that
these things are not elements of language.

In operator grammar, these seeming elements are seen to be byproducts of
relations and processes involving simpler elements.  Most perspectives
agree that these simpler elements--e.g. words, morphemes, sequences of
them-- are elements of language.  The relations are relations of word
dependency (and word classes based on these dependencies), and the
processes involve assertion ("predication") plus reduction of the
actually pronounced form of a word when the gain on control for its
being pronounced is reduced.  The two perspectives are in general
agreement about much of this, though their ways of talking about it
differ.

There are other ways of describing language that see S, NP, etc. and as
byproducts of the simpler elements and relations that they do
countenance.  However, these other systems are like the familiar
phrase-structure-based systems in that they, too postulate ancillary
metagrammatical elements and relations as somehow basic.

Harris's aim was a "least grammar", since any additional objects and
relations can only obscure the objects and relations by which language
"carries" information.

    Bruce
    [e-mail]