[Corpora-List] No poverty of the stimulus

[Michael Maxwell]

I could try to reply to all the disagreements that people have had with my
views, but that would make for a lot more verbiage than this list will
tolerate (and take more time than I ought to spend).  I think Geoffrey
Sampson's response (not solely directed to me) is perhaps the clearest and
best laid out, so let me take that as a point of departure.

> But my main point is that, _even if we
> accept the claim that the child's data contains only positive and no
> negative information_, it CANNOT be the case that this makes it
> logically impossible to infer a grammar (that is a general theory using
> a limited range of principles to account for the numerous individual
> observed instances), because the natural sciences routinely produce
> general theories to account for empirical observations, and we know that
> natural scientists _never_ observe events violating physical laws.

There's at least one step in this syllogism that I don't follow.  Namely,
how do we know that learning (or inferring) a grammar is the same kind of
task as learning a natural language?  Granted, in both cases we're trying
to reduce a large set of observations to some general laws; but the
question is whether those general laws (the end goal of the analysis) have
the same character in the two cases.  I rather think they don't, on both
qualitative and quantitative grounds.

One qualitative difference is the role of exceptions.  Exceptions to
generalizations abound in languages--indeed, the entire approach of
Optimality Theory is based on defeasible constraints.  (I'm not a
proponent of OT, but it is useful in this context for making the role of
exceptions to generalizations clear.)  While the natural world abounds
with *apparent* exceptions to natural laws, the laws of physics in fact
don't have exceptions.  So it seems to me that there is a strong
difference here in the laws which govern the two domains, and
therefore--it seems to me--one cannot a priori say that language learning
does not require negative evidence just because natural science (in
particular physics, perhaps the clearest case) doesn't.  *Whether*
language learning in fact requires negative evidence is an empirical
question, not decidable on the basis of analogy with natural science.

(There is also a question of whether controls in scientific experiments
are in fact a way of bringing negative evidence to bear on questions.  And
of course John Goldsmith has pointed out that particle physicists
routinely use a sort of negative evidence.  So I think it's questionable
whether scientists don't in fact rely on negative evidence.)

> Mike Maxwell then changed the subject, to my mind, by arguing that
> natural-language grammars are so much more complex than the laws of
> physics that theory-inferring techniques which might work in the natural
> sciences would surely be inadequate for formulating language grammars.
> We could argue about this, but it is a quite separate issue from what is
> normally called the "poverty of stimulus" argument.  This latter point
> of Maxwell's is a quantitative point -- inferring theories from
> positive-only data might get a certain way but can't get far _enough_ to
> cope with the complexity of English, Portuguese, Chinese, etc.  To
> pursue that we would presumably have to find ways of putting numbers on
> how complex the theories are which could be derived in a given period of
> time from positive-only data, how complex natural languages are, how
> much time is available to the child, and so on.

One can give numbers to quantify the question of "how complex natural
languages are"; John Goldsmith's work on morphology learning does exactly
that.  But just to give an intuitive idea of the difference, Newton's laws
of motion can be given in three short paragraphs (see e.g.
http://en.wikipedia.org/wiki/Newtons_laws); grammars of natural languages
are book-length (and if you've ever tried to implement such a grammar
computationally, you know that even this is insufficient).  Hence my
statement in an earlier post that natural language grammars are orders of
magnitude more complex than Newton's laws.

> But the "poverty of
> data" argument is not about numbers at all, it is an absolute argument
> which claims that there is a _logical_ impossibility about the concept
> of formulating a grammar without prior knowledge based exclusively on
> positive instances.

You bring out a good point; the complexity must be qualitative, not just
quantitative, for the poverty of data argument to go through.  I would
argue that it is both, although I have not made much of the qualitative
argument here.  To do so might take a book-length response (like your own
book).  I am however arguing that it is not at all clear (in fact,
probably false) that learning a language is the same as learning Newton's
laws (e.g. in the different role of exceptions in the two domains); and
therefore simply claiming that because discovering laws of physics doesn't
require negative evidence, acquiring a language can't require negative
evidence either, is not a valid argument by itself.  Languages are far
more complex.  They're obviously more complex in quantitative terms, and
also more complex in qualitative terms (such as having exceptions) than
physics.  Whether the qualitative differences are such as to require
negative evidence is an empirical question, not one that can be argued by
a putative analogy between physics and language.

   Mike Maxwell
   CASL/ U MD


_______________________________________________
Corpora mailing list
Corpora at uib.no
http://mailman.uib.no/listinfo/corpora