Feasible Learnability as a Universal Constraint on All, even Alien, Languages

This is an ObConLang - type post.

It's a fact that every human natlang must be learnable by an infant
who knows no language at all.  The infant must be able, moreover, to
learn it in a reasonable amount of time with a reasonable amount of
effort from a reasonable amount of data.  The data in question will
not be organized with more than reasonable effort on the part of the
infant's instructors.

This fact actually puts non-trivial constraints on the class of all
possible human natlangs.  Knowing these constraints, in turn, aids
the learner in learning the language from the sparse data available.

The same is true when an adult learns an L2.

It is a really, really safe bet that no human infant is conscious of
taking advantage of these constraints when acquiring his or her L1.
Probably, few adults are conscious of them when acquiring L2s,
either; and, in case I am wrong, and more than a few are so aware,
it's a safe bet that they are aware briefly, fleetingly, seldomly,
and probably belatedly.

However, when two species which are alien to each other make official
first contact, they will try to arrange that they are represented by
persons who are linguistically sophisticated for purposes of learning
one another's languages and teaching each other one another's
languages.  Chances are that hardship will prevent at most one of
them from being so represented.  Thus, at least one of the two
representatives who are trying to acquire each other's language, and
help the other acquire his/her/its/their own, will be fully aware of
the constraints imposed by "feasible learnability" and able to take
advantage of these constraints to speed up the learning.

----------

Among the things which help language learnability:
* The learner may not be required ever to settle on just one grammar;
and/or may not be required to find "the" "correct" one.  Instead, the
learner may be required to find one which
is "probably" "approximately" correct, maximizing the "probability"
and/or improving the "approximation" as much as possible, given the
available data.  How this is done; There is assumed to be a certain
probability distribution over the utterances.  For given tolerances
epsilon and delta, we want the total measure of the set of actually-
grammatical utterances which the learner incorrectly would
hypothesize are ungrammatical, plus the measure of the set of
actually-ungrammatical utterances which the learner incorrectly would
hypothesize are grammatical, to be less than epsilon -- that's
the "approximately" part -- with confidence greater than one-minus-
delta -- that's the "probably" part.
* The learner may be presented with negative data as well as with
positive data.  That is, sometimes the learner may be presented with
an ungrammatical utterance and informed that it is ungrammatical,
instead of always and only being presented with grammatical
utterances and informed that they are grammatical.
* (This happens even with babies learning L1s) The learner may
(sometimes) be presented with an utterance together with its meaning,
or some hint as to its meaning.
* (This happens even with babies learning L1s) The learner may
(somethimes) be presented with an utterance together with some help
as to its phrasing, for instance, at the most, an unlabeled parse
tree, which will show exactly what gets grouped with what -- what the
phrases are, but not what types of phrases they are.
* The learner may be able to produce utterances and find out if the
produced utterances are or are not grammatical.
* The learner may be able to produce utterances and find out whether
or not they mean what the learner hoped they meant.
* The learner may be able to assign candidate meanings to utterances
and find out whether or not they are correct.
* The learner may be able to assign candidate phrasings to utterances
and find out whether or not they are correct.

----------

Has anyone else heard of:

Ranked Node Rewriting Grammars?
Stochastic Ranked Node Rewriting Grammars?
Tree Adjunct Grammars?
Tree Adjoining Grammars?
K-Local Grammars?

----------

There seems to be some sort of idea that a 3-local Ranked Node
Rewriting Grammar of rank <=3 with sentence-depth 2  (that is,
sentences can have sentences as participants that, in turn, can have
sentences as participants; but these latter sentences cannot have
sentences as participants, in their turn.) would be polynomially-
learnable and this would be equivalent to a Minimalist Program
grammar.  No, I don't know what all that means.

Apparently computational biochemists have gone further with this than
anyone else.  Weirdly, the programs it takes to get a computer to
acquire an L1 the way a baby does, are just like the programs it
takes to get a computer to look at the sequence of amino-acids making
up a protein, and predict how it will fold.  Since mis-folded
proteins are behind Alzheimer's disease, scrapie, bovine spongiform
encephalitis, Creuzfeld-Jakobs' disease, and kuru, this is a hot
topic.

----------

A good story would be:

Two linguists -- a human and a non-human -- are assigned to teach
each other one each of their own languages, and learn each one of the
other's languages.  The human, at least, assumes the alien's language
is feasibly-learnable, and takes advantage of the constraints that
imposes to help speed up the learning.

However, doubts set in.

These doubts are externalized by (one or some of) the human
linguist's superiors.

One of two second twists can occur:
Victory of the universals; the original assumption was right, and the
human and alien linguist both win through and learn each other's
languages quickly, by assuming it can be done.
Victory of the weirdness of the universe; it turns out that "feasibly
learnable" means something very, very different for these non-humans
than for humans, or even than for human-made computers; they have a
native U.G. that's just plain strange.  (Maybe they'r Yksmohcites.)
The linguists learn each others' languages, but it takes them an
awfully long time; or, perhaps, they invent a lingua franca between
the two of them faster than either of them learns the other's
language.

-----

What does anyone think?

Any comment welcome.

Tom H.C. in MI









		
---------------------------------
 Yahoo! FareChase - Search multiple travel sites in one click.