Re: Chomsky vs. Skinner on Language

From: Stevan Harnad (harnad@cogsci.soton.ac.uk)
Date: Sat Mar 02 1996 - 19:26:25 GMT


> From: "Chatwin Judy" <JAC295@psy.soton.ac.uk
> Date: Tue, 27 Feb 1996 09:40:22 GMT
>
> sh> Not quite. He said that what they DO do has been (largely) shaped by
> sh> reward much the way a rat is shaped to press a lever for food.
>
> Does this mean that reward is more influential than punishment?

More effective in shaping behaviour. Punishment may inhibit something,
but it can't shape it positively.

> sh> (This does not rule out doing something for no particular reason, or
> sh> because you were born that way. It's just that most behaviour is the
> sh> result of a prior history of rewards.
>
> Is this as a result of our own experiences and/or the experiences of
> others - if these experiences are conflicting then which one takes
> priority? (This may not only refer to language.)

Both; others may reward us based on their own history of rewards. But
they may also instruct us: tell us. But to do that, we already need the
language...

> sh> Skinner would have said that, whatever those rules might be, the child
> sh> must have learned them by having been rewarded for saying things
> sh> grammatically and not rewarded (say, ignored) for saying things
> sh> ungrammatically. Then gradually, like the rat learning to press the
> sh> lever, the child learns to produce only the grammatical strings of words
> sh> and not the ungrammatical ones, according to Skinner.
>
> But surely there are some occasions when meaning/desires can be
> communicated without correct grammar - there is enough information
> transmitted to 'get the message across' even though there may be some
> grammatical errors.

True, but irrelevant, since the question here is: "How do we learn to
speak grammatically?" Not "How do we learn by means other than language?
or "How do we understand language even when it's not grammatical?"

> sh> True, but this was Chomsky's CONCLUSION from the evidence, not just an
> sh> assumption he made. It turns out that grammar HAS to be innate, because
> sh> the trial and error evidence the child gets -- i.e., everything he says
> sh> and hears till age 4 -- is insufficient for the child (or anyone) to
> sh> derive the rules from it by trial and error.
>
> Is there concrete evidence which says that a child never makes
> grammatical errors - surely it could be argued that they don't
> continue to put together certain constructions because, although the
> meaning may be successfully communicated, those constructions are not
> repeated and heard by the child?

The child makes trivial grammatical errors ("I holded"), which ARE
corrected by feedback, but it doesn't make errors in Universal Grammar
(UG): How does it KNOW no to make those errors?

> The example was that you can say "Who did he
> sh> hit that went out?" but you can't say "Who did he think that went out?"
> sh> gives you the flavour of the kind of abstract rule we all "know," yet
> sh> we could never have learned from trial and error (because no one ever
> sh> says or hears "who did he think that went out" -- except in a
> sh> linguistics seminar on Universal Grammar!)
>
> What would happen if a construction such as this one was regularly
> used - would it be accepted? As I understand it, if Chomsky's
> argument holds firm then this would/could never be part of English
> grammar. How can it be argued across different languages when
> constructions differ (please be aware that I do not have any great
> knowledge of other languages, but would welcome comments on how this
> applies universally)?

Any PARTICULAR violation of UG could of course be accepted as a special
case; that's not interesting. But if all or most of the feedback a child
got were violations of UG, the prediction is the child could not learn a
language at all.

> sh> (By the way "Who did he hit that went out?" is also ungrammatical in the
> sh> old, non-Chomsky way. It should be "Whom did he hit that went out?" But
> sh> that's not the kind of grammatical rule at issue here, because that rule
> sh> CAN be (and is) learned by trial and error and instruction and
> sh> example.)
>
> Am I right in thinking that there are some types of grammatical rules
> that fall into the title of 'Universal Grammar' and some that do not,
> please could someone clarify this for me?

You are right. And in general, we know the non-UG rules as the "grammar
of English" (don't say "ain't", etc.). Those are NOT the rules at issue
here. The ones at issue are highly abstract ones, that linguists have to
work hard to infer -- not rules that every schoolteacher knows and
teaches. Those UG rules are not taught (school-teachers don't know what
they are: linguists are still trying to work them out!), yet the child
never violates them. Hence he must already "know" them in some
unconscious way at birth.

> sh> It is the stimulus that is impoverished, not the grammar.
>
> I am still not very clear on the meaning of the 'Poverty of the
> Stimulus' - please could someone explain it for me.

What the child hears, and the feedback the child gets for what he says.
Compare grammar rules and pronunciation rules: The child hears things
pronounced correctly, and gets feedback from pronouncing them wrong. The
combination of examples of imitation and corrections for errors made is
the "stimulus." In the case of pronunciation, this set of examples IS
rich enough, informative enough, to arrive at the correct
pronunciations; in the case of grammar it is not.

Here's another example:

Supposing I was trying to tech you the rule: "even numbers". I tell you
to start saying numbers, any numbers. Whenever you say an even one, I
say "correct," whenever you say an odd one I say "incorrect." After a
while, by this purely Skinnerian feedback, you would know what even
numbers were correct.

This of course assumes that you already know what numbers are and even
what even numbers are, but you don't yet know what is "correct" in this
situation; I could have used "prime numbers" as my example; that would
be harder, and would take longer, but again, the examples are rich
enough so you could learn the rule.

But if, for example, I told you to say letters instead of numbers, and I
gave you feedback for whether a letter happened to be the first letter
of a prime number when pronounced in any European language, then chances
are that that "stimulus" would be too impoverished for you ever to learn
the rule.

That's the situation the child is in with "learning" UG, yet he does
"learn" it: How come? Because he already has it when he is born, so he
does not make the mistakes, hence need not be corrected. The only
examples he produces or hears are "positive" ones (as if all he heard
or said were prime numbers from the very beginning.

> sh> The strange case is the one where the rule itself cannot be learned from
> sh> examples (because the examples are too impoverished, they are missing the
> sh> "negative evidence" from getting the rule wrong, and having it
> sh> corrected).
>
> I think this helps clarify some of my understanding regarding the
> Poverty of the Stimulus.
>
> sh> Remember it wasn't a definition but a criterion (sort of a test for
> sh> whether something is or is not a language): There seem to be no
> sh> "primitive" languages. Language seems to be all-or-none: If you can say
> sh> something in it, you can say anything in it.
>
> Does this connect to the 2,000 root words that you mentioned?
> If you think otherwise, tell me.

Yes it connects: Not only can you say anything in any language that you
can say in English; not only can you intertranslate any language with
any other language; but you can also say anything you can say in English
(or any language) in a relatively small fragment of English -- say, the
~2000 words in that basic English vocabulary dictionary, from which all
the other words are defined.

> Looking at my notes from last week I see that I see that I have
> bottom-up processing and top-down processing mentioned but please could
> somebody relate it to the understanding of the learning of language. (I
> remain convinced that my memory fades with age!)

Bottom-up means starting from concrete examples and abstracting rules
from them. Top-down means applying rules you already have to concrete
examples. It's a bit related to induction vs. deduction.

> I also have written down that if Skinner's theory is to be believed,
> then what about abilities such as pattern recognition?

Use that in your Explaining the Mind essay...

> I have read Stevan's E-mail to us regarding the vanishing
> intersections and although I thought I understood it, I do not think
> I do - again can anybody explain it to this 'kid-sister' and use your
> imagination regarding the age!!

One you take a bunch of circles and overlap them, the intersections are
the parts that they all have in common. If they don't all have a part in
common, the intersection is empty, it has vanished.

That's just a picture. It applies to sets of things. Think of sets of
things having members. The intersection of all the people who are
Psychology Students in the UK, Southampton Undergraduates, and Females
is: female, Soton Psychology UGs. That intersection is NOT empty. It
does not vanish. The intersection of all the people who are Psychology
Students in the UK, Southampton Undergraduates, Female, and under the
age of 3, on the other hand, is, as far as I know, empty.

The vanishing intersections argument against bottom-up learning of
abstract categories from sensory examples is this: Maybe we know what an
"apple" is from what it is that all the apples we have seen have in
common: the intersection of all the features of all the apples we've
seen. And that non-empty intersection can be learned by trial and error
from experience. But something more abstract, like fruit, or food, or
object, or entity could not be learnt bottom-up from experience, because
the intersection of all the cases -- say, all the things you've seen
that are food -- would be empty: They have nothing in common.

So (say those like Jerry Fodor, who regard this as a kind of a "poverty
of the stimulus" argument for abstract concepts, analogous to the one
for the unlearnability of UG), the concepts must be inborn.

Do you agree with this argument?

Chrs, S



This archive was generated by hypermail 2b30 : Tue Feb 13 2001 - 16:24:15 GMT