Re: principles of collective organization
Rob,
Just some historical points:
> American Structuralism was essentially _all_ distributional.
> Chomsky killed it in the '50s by pointing out it led to
> inconsistent representations (=no primitives). Now that people
> have forgotten Chomsky's objection distributional methods
> have become a popular machine learning technique, but they
> are searching for primitives again.
The term I would prefer is _empirical_ instead of either
_distributional_ or _primitives_. Bloomfield's first book,
Bloomfield, Leonard (1914) An Introduction to Language,
reprinted by J. Benjamins, Amsterdam, 1983.
was very strongly influenced by the semantic theories
of German psychologists such as Wundt. But in the 1920s,
he was seduced by the behaviorists, who persuaded him that
anything that could not be directly related to observable
data was "unscientific". That led him to the strongly
data-oriented, anti-semantic, behaviorist structuralism
of his second book:
Bloomfield, Leonard (1933) Language, Holt, Rinehart,
& Winston, New York.
That book was the primary introduction to linguistics for
the middle of the 20th century. A very popular structuralist
goal was the search for "grammar discovery procedures", which
would start with large volumes of linguistic examples and
derive the grammar that characterized them. Various studies
along those lines used finite-state machines and statistics
as methods of analyzing and characterizing the grammars.
In his 1957 book, _Syntactic Structures_, Chomsky retained
Bloomfield's anti-semantic bias, but he fought against C. C. Fries
(who did a very nice derivation of grammatical categories from
a corpus of utterances in the early 1950s) and Claude Shannon's
information theory (which was very popular among psychologists
and linguists). Chomsky claimed that the data an infant received
was insufficient to derive a complex grammar and that most of
an infant's linguistic mechanisms had to be innate. In his work,
Chomsky ignored real data and began the practice of using a native
speaker's intuitions (i.e., made-up examples) as the basis for
doing linguistics.
The trend toward applying statistics to large volumes of data
really began in earnest when computer systems began to have
disks that were large enough to store significant volumes of
data. But I think that the trend toward using real data instead
of made-up examples is more significant than the idea of using
statistical methods.
As for primitives, they could be derived from data or from
intuition, and they could be syntactic or semantic. I believe
that the option of ignoring or searching for primitives is
orthogonal to the issue of an empirical or arm-chair approach.
John