[matrix] Pronoun Relations

[Ann Copestake]

I'm not sure what got minuted, but I do clearly remember a discussion about
pronoun relations in Norway ...

Currently, what we get out of having the relation rather than just an
index is the cfrom/cto information which allows us to take cataphora
into account.  As far as I can see, it would be tricky to reconstruct
this if we just have an MRS with indices and no relations for the
pronouns.  For instance, the MRSs for A and A' below are presumably
identical (modulo cfrom/cto and any information structure) but we need
to be able to distinguish them.

A: Because the cat purrs, Kim likes it.
A': * Kim likes it, because the cat purrs.

I have no clue about what happens for Japanese, but we'll have a more
poorly-performing algorithm for English if we don't take this into
account.  Maybe one should do this off the cfrom/cto values for the
relations for which the pronouns are arguments, but the Lappin and
Leass treatment, for instance, just involves giving a negative weight
to pronouns coming before their putative antecedants.  

Obviously this is a bit of a hack, but I would argue that it's
currently a useful surrogate for a proper approach to information
structure.  There is also a practical point that concerns evaluation
of anaphora resolution: we need to be able to locate the pronouns in
the text and it's a pain to do this on the basis of indices alone.  In
fact, I think it may be impossible in the limit - `Kim gave it it' vs
`Kim gave it to it' - silly example, but the point is it would entail
non-trivial code writing to do this properly.

Anaphora resolution approaches like Lappin and Leass have
reasonably straightforward approximations in MRS (using the ARG
numbering instead of the SUBJ>OBJ etc weighting).  It would be
interesting to see whether there's any difference in performance.

Incidentally, I think that there's an argument that we should be
distinguishing between the different pronouns in terms of their
relations, since right now, we're only capturing distinctions that
make it onto indices.  This implies, I think, that we'd have to have a
politeness value on indices to record the German du/ihr/Sie
distinction and presumably some sort of dialect feature to indicate
non-standard pronouns (e.g., the `us' for 1sg in some BrE dialects).
I don't know what sort of variation is possible, and maybe it's OK to
say that all variations in pronoun use are either ignored or put on
indices, but I do worry that this could lead to index proliferation as
new languages are dealt with if we go down the route of removing
pronouns.

I think there are existing mechanisms to allow a clean transfer
approach.  I believe that there is the possibility of monolingual
rules in the transfer machinery.  So you write a rule that just drops
all pronoun relations from languages like English before transfer.
You need another rule to introduce them for generation, but this is
very similar to what you'd have to write anyway for the null semantics
items in the generator.  I believe that this is the right mechanism to
use for investigating systematic differences between languages: it's
far less disruptive than changing the Matrix.  It may well turn out
that once an inventory of rules has been developed, an overhaul of MRS
generally is called for.  But given that many of us are using it for
applications other than semantic transfer, we'll have to try to take
different needs into account.

Ann

> 
> I believe the starting point in the discussion in Norway was what to
> do about pronoun relations for *dropped* arguments.  Here's the link
> to the notes from that discussion:
> 
> http://wiki.delph-in.net/moin/FeforDroppedArguments
> 
> I'm still not clear on what information is available in the
> pronoun_rel that is not already available from the index itself (say,
> in combination with the feature PRON_TYPE).  Furthermore, I resist the
> idea that putting in pronoun relations in the case of dropped
> arguments is the right path.  In general, we need an algorithm for
> anaphora resolution which will work when there are no overt pronouns
> in languages like Japanese.
> 
> Emily
> 
> On Thu, Feb 01, 2007 at 12:22:29PM +0000, Ann Copestake wrote:
> > 
> > we talked about this in Norway.  It is inconsistent with supporting any 
> > standard approach to anaphora resolution, which is something we have to do.  
> > if you want to make this change, therefore, you have to come up with an 
> > algorithm for anaphora resolution which will work when there are no overt 
> > pronouns to resolve.
> > 
> > sfd at u.washington.edu said:
> > > What's more, having
> > >  pronoun relations in the MRSs of only some languages is troublesome for our
> > >  MT system, requiring rather elaborate transfer rules.
> > 
> > I'm afraid that once you go above very small scale systems, you'll find the 
> > need for elaborate transfer rules in many places, so this is not a persuasive 
> > argument.
> > 
> > Ann
> > 
> > 
> >