THEORY: OT Syntax
|From:||dirk elzinga <dirk.elzinga@...>|
|Date:||Tuesday, November 21, 2000, 10:13|
On Sun, 19 Nov 2000, jesse stephen bangs wrote:
> Marcus Smith sikayal:
> > That depends on your theory of "Gen". Prince and Smolensky actually provide
> > two different models. One - the one nearly everybody uses - generates all
> > possible forms, then selects the most "optimal" form based on the
> > constraints. The other only generates the forms relevant for the next
> Huh? This implies that Native Speaker X, before he says anything, stops
> to generate every single conceivable combination of words and then applies
> a bunch of constraints to rule out all of the illegal forms. That gets
> impossible pretty fast. Do I misunderstand "Gen"?
Not at all, but you misunderstand perhaps what OT is. It is not
intended to be a model of Natural Language Processing, so there is no
serious claim that speakers actually generate all possible forms. OT
is intended to be a model of linguistic competence. That is, the
output of an OT grammar should accord with native speaker intuitions
about what constitutes a grammatical utterance. This, BTW, is the
standard set up for generative grammars in general.
> > constraint on the list. After that constraint has made a decision, the
> > optimal candidate is used as the base form for generating the next set to
> > be judged from. McCarthy is the only person I know of to use such a model
> > in Phonology (but I'm not a phonologist, so I don't know the literature).
> > Under this second theory of Gen, you do not get the forms generated
> > "willynilly". Indeed, McCarthy has suggested that Phonology should follow
> > Syntax in using such a Gen. (That was at the West Coast Conference of
> > Formal Linguistics, 2000).
> This seems reasonable, at least with respect to syntax. I don't quite see
> how this would work with phonology.
It would work in the same way. You take an underlying representation,
and generate all possible forms which are "one step" away from the UR.
You then evaluate these forms according to the constraint set. The
winner is then subjected to the same process; generate all possible
candidates which are one step away from the input and run it through
the constraint set. Keep doing this until the grammar converges (until
the steps of UR-Generate-Evaluate fail to produce anything better).
I don't know McCarthy's WCCFL paper, but at NELS the year before he
showed that the serial interpretation of Gen fails, and that the
parallel form should be preferred. I recently read a paper by Matthew
Chen who uses OT to evaluate candidate *derivations;* I need to read
it again to fully understand the arguments, so I won't comment on it
> Actually, I'm still quite confused about the theoretical framework for
> OT. How does the whole process get started? Is there some sort of
> maximally unmarked phrase structure to generate forms to choose from? Are
> sentences generated at random until one comes out right? The examples of
> more complex analyses that I've seen are valid, but the basic baffle me.
This is the big question that everyone seems to be tiptoeing around in
OT syntax; what is the nature of the input? I'm not sure if anyone has
answered the question satisfactorily.