[developers] Wish for scoped MRSs in the LKB

Francisco Costa fcosta at di.fc.ul.pt
Fri May 12 13:03:29 CEST 2006

Emily M. Bender wrote:
> Does this solution scale to cases with more than two quantifiers?

I believe it might. You would be saying that everything but `a certain' 
has a BODY with SCOPE *non-widest* (and so `a certain' must outscope 
them all), perhaps with the exception of quantifiers of proper names and 
pronouns, for which you might want to assign even wider scope (and you 
would need a more sophisticated hierarchy under *scope*). You would not 
say anything about the BODY of `a certain', otherwise you can't have 
multiple instances of `a certain'. But I can't test it for obvious 
reasons, so...

But my point was not the analysis itself, but rather that it might be 
interesting if grammar writers could put constraints on handles in order 
to constrain scope, in that for me it is really intuitive that handle 
variables that can represent the same handle should be unifiable, and 
the ones that are not unifiable should not be able to stand for the same 
thing. I was not asking for this so that this particular analysis can be 
implemented, but because I think it is what you would expect to be 
happening, and so it would be a general improvement.

Francisco Costa

> Emily
> On Thu, May 11, 2006 at 01:02:03PM +0000, Francisco Costa wrote:
>>I would like to send in a wish for the LKB.
>>As far as I can tell from experimenting with it, asking for scoped MRSs 
>>does not perform unification operations on the handles that end up being 
>>I'm trying to do the following. When you have a sentence like `all men 
>>love a certain woman', the existential quantifier must outscope the 
>>universal one. So I tried this: (1) add a feature SCOPE to handles, of a 
>>type *scope* which has two subtypes *widest* and *non-widest*, (2) 
>>constrain the LBL of the `a certain' quantifier to have SCOPE *widest* 
>>and (3) constrain the BODY of the `all' quantifier to have a SCOPE 
>>They do not unify (and manual unification in the AVMs fails), so I was 
>>hoping that the reading *all_q(x1, ..., a+certain_q(x2, ..., ...))* 
>>would not come up when you ask for scoped MRSs, but it does. Using 
>>subtypes of *handle* instead (no feature SCOPE) doesn't do it either.
>>Is it possible/feasible to have an algorithm for scoped MRSs that 
>>performs unifications on handles? I think I've read somewhere that it's 
>>already exponential on the number of NPs, but I guess feature structures 
>>of type *handle* will never be very big.
>>Thank you in advance,
>>Francisco Costa

More information about the developers mailing list