[developers] `fragmented' scoping MRSs

Emily M. Bender ebender at u.washington.edu
Fri Jul 25 15:22:53 CEST 2008


Unfortunately, that's not similar from the grammar's point of view
(this is Wambaya, in case I hadn't said so yet):

I'm having the verbs contribute the quantifiers for their arguments
because of the discontinuous NPs.  In some cases, though, the
arguments are not overtly realized by anything (noun or modifier),
so I'm stuck with the quantifier alone.

Emily

On Fri, Jul 25, 2008 at 3:10 AM, Ann Copestake
<Ann.Copestake at cl.cam.ac.uk> wrote:
>
> yes - a relatively simple solution from my point of view, which might
> possibly be a reasonable one, is to not put in quantifiers and type
> those variables so they are recognised as being allowed to be free,
> and implicitly existentially bound.  cf event variables.  (This may
> involve code tweaks, but not conceptual ones ...)
>
> Ann
>
>>
>> Hi Ann,
>>
>> I'm getting unbound RSTR values precisely where I have no
>> relations for the RSTR: the argument position was filled neither
>> by a nominal head nor by a modifier.  Given the current shape
>> of the grammar, it would be very inconvenient (but I'd hesitate
>> to say impossible) to put in pronoun relations for these cases
>> (or alternatively, to not put in quantifiers, either).
>>
>> It may just be the case that this is only happening with
>> existentials.  The universal shows up as inflection (on nouns
>> and adjectives), so it brings its restriction along with it.
>> I don't know if `no' can appear without any thing else in the NP,
>> but I can at least take a look for that in the data that I do have.
>>
>> Emily
>>
>> On Thu, Jul 24, 2008 at 11:48 AM, Ann Copestake
>> <Ann.Copestake at cl.cam.ac.uk> wrote:
>> >
>> > Having a RSTR value specified by semantic composition is fairly basic
>> > to the way I've understood generalised quantifiers, but people with
>> > more expertise might be able to point at counterexamples.  More
>> > concretely, if there's an empty RSTR value, then all that the scoping
>> > code would be able to do with quantifiers would be to produce variants
>> > with different distributions of the EPs between the RSTR and the BODY.
>> > e.g., it would be like saying there's a language where you can have
>> > {teach sleep every} and get the readings `every teacher sleeps' and
>> > `every sleeper teaches' (or whatever).  Would this be the desired
>> > behaviour?  If it only happens with cases where the RSTR/SCOPE
>> > distinction is irrelevant to the models (like `some' under the
>> > non-zero set interesection interpretation), this would also be an
>> > interesting finding.
>> >
>> > Best,
>> >
>> > Ann
>> >
>> >> Thanks, Ann.  I did notice that the scoping code requires the
>> >> RSTR value of a quantifier to be related to something (not
>> >> currently the case always in the Wambaya grammar...).  Is
>> >> this a deep requirement?  That is, is it bad to have an empty
>> >> RSTR value?
>> >>
>> >> Emily
>> >>
>> >> On Fri, Jul 18, 2008 at 6:18 AM, Ann Copestake
>> >> <Ann.Copestake at cl.cam.ac.uk> wrote:
>> >> >
>> >> > yes, I've seen this when playing around with grammars.  (I would also expect
>> >> > to see non-frag MRSs that don't scope.) I believe the test for fragmentation
>> >> > is checking for explicit label/hole relationships between all the EPs (either
>> >> > label equality or qeq), though I haven't seen a spec, so I might be wrong
>> >> > about this.  The scoping code in the LKB doesn't require that qeqs are there -
>> >> > it just ensures they are obeyed if they are present.
>> >> >
>> >> > Ann
>> >> >
>> >> >
>> >> >
>> >
>> >
>
>



More information about the developers mailing list