[developers] Zulu as an interesting case of generator optimization

[Emily M. Bender]

Dear all,

I've turned up something interesting in using the generator
with the Zulu grammar (developed for my class by Kelly O'Hara).
Zulu inflects verbs for both subject and object concord, and
we are modeling this with a feature NCL (analogous to GEND)
on referential indices.  

Since there are 18 noun classes and transitive verbs inflect
(potentially) for both subject and object agreement, this
leads to a lot of generator edges!  I was actually surprised to
see this, as the inflection does have semantic effects --- it
doesn't add a rel, but it does change the value of a feature
inside one of the rels (ARG1/ARG2 on the verb's relation).
And then these edges can of course feed various other rules
(such as the head-opt-subj rule), so generation for this language
is pretty slow.

Is there some way to have the generator filter lexical rule
applications for this sort of case?  If not, what difficulties
do you see (in principle) with implementing this kind of a check?

Thanks,
Emily