Re: Universal grammar

In article <1162112325.958675.203450@xxxxxxxxxxxxxxxxxxxxxxxxxxxx>, "Rob
Freeman" <groups@xxxxxxxxxxxxxxxxxxx> wrote:

Back to language. What do you think of Rob Freeman's
approach of a grammr based on examples instead of rules?
I find it promising. That is the way I proceed in English.
I don't consult grammar books, I rely on songlines and
sentences I remember.

When I learned language, one focused much on grammar and spelling, but
nowadays, one has realized that it is important to practice, practice and
practice. The same as always has been known in say music performance.

It is easy to slip into thinking this is a learning issue. That is
unfortunate, because with the assumption that we are talking
principally about learning, comes the assumption there is something to
learn (usually rules.) The "central simplifying generalization" I work
with is not so much that we learn language from examples. It is that
the examples themselves are what we learn, that the examples themselves
are a particularly compact representation for a language (the most.)

The problem is that as soon any data enters the human brain, it is
immediately processed by the higher cognitive authorities. Then what has
been processed is what is memorized. Research about autistic savants (or
idiot savants)
  http://en.wikipedia.org/wiki/Autistic_savant
support this idea. These people have damage in some of the higher
cognitive filtering areas of the brain, and by that, they are able
to truly learn examples as they are. I feel sure this is not what you mean
or intend by learning.

That said, there can be a feeling of "learning by learning rules" when
learning a closely related language. To the extent that works, I think
it works only because the languages are closely related.

So, as you work through big amounts of examples, your brain will start
to intuitively develop rules, of which no have no formal knowledge of.

By contrast, computers do not have any higher cognition filtering system.
So they only get the rules that the programmer explicitly puts into them.
Therefore, they do not have the capacity to learn by examples alone. AI
might provide such a filtering system, though still quite primitive from
the point of view of the human brain, but a human then needs to program
that filtering system.

According to this theory a "rule" is a name for a (sub-)set of
examples. When, for historical reasons, the sets are largely the same,
it is possible to quote a rule and simultaneously "learn" (actually
analogize in bulk from your own language) all the forms in that set.

When I use the word "rule" in this context, it is just a collective
designating any type of higher cognitive thinking. One needs to be careful
when distinguishing such an intuitive concept with any type of
formalization within a theory.

It is perhaps easier to see that analogous sets are what grammatical
labels give you when the sets are not strictly grammatical. So for
example, a speaker of one European language learning another just needs
to learn the word for, say, "that" (which usually functions equally as
a determiner, a pronoun, a subordinating conjunction, and the basis of
articles when they exist...), "which" (also functions commonly as all
of a determiner, a pronoun, and a subordinating conjunction across
languages), "and" (which always lets you both list and conjoin to build
arbitrarily complex phrases), and he or she already has bones for the
language. These are lexical regularities, but the principle is the same
for the structural regularities we think of as grammar. For mainly
historical reasons labels like tense and mood in most European
languages, e.g. the conditional, define sets in one language which
correspond to strictly analogous sets (or paradigms) in another. What
you are doing is importing, in bulk, whole sets of analogous examples.
For a while you have to do this consciously, example-by-example, then
you feel you are "talking by translating", but after a while you
remember the examples themselves, then you are "thinking in the
language."

When analyzing this learning process, one has to carefully distinguished
between intuitive and formalized rules. We do know know of a good way to
read off a human brain with its internal intuitive rules, and extract
formalized rules from that. The human brain acts as a black box, whose
inputs and outputs the formalized rules describes.

It doesn't work when the sets are not the same. When learning a truly
unrelated language, not only does learning a rule not allow you to
import a set of examples in bulk. Without the corresponding set of
examples it is not even possible to conceive the rule (much as those
with no experience of a "tone" language find it difficult to conceive
"tone".) Someone can tell you that to build a well formed Japanese
sentence you need to decline conjoined clauses as adjectives, and
decline the whole, like a word, for courtesy, but without a few
examples that does not "mean" much.

It just means that the intuitive rules needed for learning the language
wasn't developed at early childhood. It might even mean that a grown-up
human without the right talents might be unable to learn such intuitive
rules, and must resort to formalized rules.

The more languages you learn, the easier it becomes, because you have
more "sets" to refer to by analogy.

So it probably just means that one already has more rule patterns to
invoke, thereby making the processing of examples more efficient.

So, those are a few learning implications of the model. But as I say,
they are not central.

What is central is the complexity argument, which says examples are the
most efficient way to store all the grammatical generalizations you can
make about a language.

The human brain simply does not store information that way. It stores the
processed information, not the examples alone as mere data.

"There are more patterns in text than can be captured by any one
summarization."

So you are just saying what I say above, that humans can freely mix
cognition on several levels of abstraction. You do not get abstraction by
examples alone: some higher cognitive processing is needed.

--
Hans Aberg
.

Relevant Pages

  • Re: Strong AI Thesis (No Chinese room, I promise)
    ... with the idea of a generic learning system at work. ... make vision work better, and language work better, etc. ... Logic, mathematics, and scientific method ology have ... recognized the limitations which linguistic practices im pose on ...
    (comp.ai.philosophy)
  • Re: Gradual Learning, not Reinforcement Learning
    ... You seem to believe that "learning" means acquiring behaviours the system has never exhibited before. ... You use learning HORSE as an example, and posit HOXEL as an intermediate stage, with HORSE as an "ideal" to be achieved. ... Firstly, the system must be capable of "producing language", else it cannot learn the semantic difference between HORSE and HOTEL, both of which are easy consequences of HOXEL. ...
    (comp.ai.philosophy)
  • The Gaelic Society of Moscow speaks out!
    ... TEN REASONS FOR LEARNING IT ... ©Gaelic Society of Moscow,2003 ... language if you have few opportunities to use it in addition, ...
    (soc.culture.scottish)
  • Re: The Gaelic Society of Moscow speaks out!
    ... TEN REASONS FOR LEARNING IT ... ©Gaelic Society of Moscow,2003 ... language if you have few opportunities to use it in addition, ...
    (soc.culture.scottish)
  • RE: C# book for 16 yo?
    ... First I would suggest first learning Visual Studio.NET. ... Then I suggest 2 books to actually learn C# - Microsoft has a book called ... important to quality programming. ... language that combines all the best features of previous languages and is ...
    (microsoft.public.dotnet.languages.csharp)