[Ligncse256] Reference?

[Roger Levy]

Ben Cipollini wrote:
> Hey all,
>  
> As it turns out, it's actually theories in reference in the philosophy 
> of language (e.g. Kripke & others) that pissed me off enough to get me 
> interested in cognitive science.  It's been over 12 years since I wrote 
> a paper about how "Everest" does not refer to "Gaurisanker" (damn it!)  
> Anyway, point being: issues in reference really ... BUG ME haha
>  
> In class, Roger mentioned that some noun phrases do not refer ("every 
> dancer"; "no dancer").  I stayed silent in the lecture (I did!) , hoping 
> that by the end that this view, while not "theoretically" desirable for 
> me, might be computationally desirable.  This email is my argument that 
> it's neither theoretically desirable nor computationally desirable.
>  
> Theoretically:
> To say that "every dancer" does not refer is to say that any noun in a 
> work of fiction does not refer.  Well, that's crazy, I say!  You may 
> even want to push further, to say that words with uncertain reference 
> don't refer.  But language is a type of noisy-channel communication; 
> there is no certainty in any communcation.  Where do you stop?  Instead 
> of trying to answer the question, maybe take a cognitive perspective.
>  
> Jeff Elman has suggested that meaning is a "walk in a state-space" or 
> something along the lines.  In other words, meaning is a path in a 
> computational space.  Why not understand reference as refering to 
> cognitive states, or paths in a computational space?  That would allow 
> all nouns to refer, including categories ("dogs"), quantified categories 
> ("every/some/no dancer"), fictional places & names, historical figures, 
> etc. 
>  
> Computationally:
> Well, that's fine, but as with everything computatational, I'm always 
> willing to sell my theoretical soul to the computataionl devil if I can 
> gain in performance or tractibility!  But ... as far as I could tell, 
> making the division between referring and non-referring nouns actually 
> added a layer of complexity into models.  We have to classify based on 
> referring / non-referring noun class first, then go on to determine 
> reference and reference dependence relations.
>  
> Why make the distinction?  Every dancer touched her toes; why not let 
> "her" refer to "every dancer"?  What problem does that cause?  Even if 
> you don't accept what I said above about the theoretical "goodness" (are 
> you crazy??), how would this help us computationally?
>  
> Seriously though, I'm happy to hear answers.  If you're not interested 
> in a .... "thoughtfully aggressive" reply .... you may want to let me 
> know when you answer :D

Hi Ben,

These are very interesting and important issues you're raising!

First, I would personally argue that whether linguistic reference 
directly implicates cognitive states may to a considerable extent be 
orthogonal to the question of what the mathematical structure of 
referential meaning is like.  (These could be questions at different 
levels of analysis.)  To me, the claim that meaning is a "walk in a 
state-space" is crucially a claim about the nature of incremental 
interpretation (one that I agree with: every extra bit of the linguistic 
input you process changes your beliefs about what the utterance is 
likely to mean), and leaves quite wide open the question of what the 
structure of the state-space is.

Second, let me try to lay out some of the problems that quickly arise if 
you try to make quantified NPs referential.  For an NP to be 
referential, it should be able to be mapped to an individual or group of 
individuals in the (comprehender's cognitive) "world", such that the 
truth conditions of "NP is Y" are that every member of the group mapped 
to by NP is in the set mapped to by Y.  Now for the case of "every 
dancer", it might seem that you can get away with mapping the NP to the 
group of all dancers.

There are two immediate problems that arise here.  First, this doesn't 
work for "No N" NPs.  For example:

   (1) No woman has walked on Mars.

There is no way of choosing a group of individuals G in the world for 
"No woman" to map to such that (1) means that G has walked on Mars. 
[You can't choose, for example, the group of individuals who are not women.]

Second, if "Every N" maps to the complete group of Ns, then sentence 
negation becomes problematic:

   (2) Every dancer did not twist her knee.

This sentence has two scopings, one of which would be consistent with a 
mapping of "Every dancer" to the group of all dancers (i.e., the scoping 
where no dancer twisted her knee).  But the other scoping, in which at 
least one dancer didn't twist her knee, cannot be obtained by having 
"Every dancer" map to a fixed group of individuals.  (If you don't like 
example (1), think of "all that glitters is not gold".)

These difficulties arise in both theoretical and computational/practical 
contexts.

Now, I *do* agree with you that this adds an extra layer of complexity 
into coreference-resolution systems.  But I actually think the real case 
is that coreference systems tend to drastically oversimplify the problem.

FWIW.

Roger

-- 

Roger Levy                      Email: rlevy at ucsd.edu
Assistant Professor             Phone: 858-534-7219
Department of Linguistics       Fax:   858-534-4789
UC San Diego                    Web:   http://ling.ucsd.edu/~rlevy