[Corpora-List] Looking for linguistic principles

[Rob Freeman]

John,

I'll let you and Mike thrash out whether it is "right" that American 
Structuralists were concerned with discovery procedures, and whether it is 
"true" American structuralists were all behaviourists.

Perhaps with the *distributional* aside that linguists may be the only class 
in nature which distinguishes itself by differing on all points :-)

On Sunday 16 October 2005 00:16, John Goldsmith wrote:
> I was responding to the question about the history of the subject -
> I think that was what the original query concerned: what had been said (by
> Halle, by Chomsky), and whether those early arguments had been answered or
> simply ignored.

Not to detract from the idea that generativism might rightly be seen as the 
cause rather than the result of the re-evaluation which occurred.

I was just most interested in hearing of new developments. More particularly I 
was interested in hearing about any discussion of this old debate in machine 
learning circles. Modern machine learning does not accept that language 
cannot be learned, which was the generativist conclusion 50 years ago, but it 
does not present any alternative solutions to the flaws which were found in 
the old distributional "discovery procedures", either.

I've talked to a few people doing machine learning, and none of them even seem 
aware of the problem.

It seems to be truly a "forget history, doomed to repeat it" scenario.

> So the first thing to be clear on is that nobody argued that structuralist
> distributional techniques led to inconsistent results....
>
> ...
>
> If the discussion is going to turn from one of history to one debating
> where we should go now...
>
> ...
>
> Yes, I do think an evaluation metric is critical for all theories of
> phonology, and I think that something along the lines of Minimum
> Description Length analysis is the right direction to go: the two terms of
> an MDL objective function correspond to grammar complexity and goodness of
> grammar-to-data fit, two notions that should be critical to any linguistic
> scientist --

Why do you need an evaluation metric if you are not selecting between multiple 
inconsistent results? If they were not multiple you would not need to select. 
If they were consistent you could take any one.

-Rob