Separating the questions

[Emily M. Bender]

----- Forwarded message from Mail Delivery Subsystem <MAILER-DAEMON at siamang.ling.washington.edu> -----

Date: Fri, 25 Jun 2004 13:42:33 -0700
From: Mail Delivery Subsystem <MAILER-DAEMON at siamang.ling.washington.edu>
To: <bender at siamang.ling.washington.edu>
Subject: Returned mail: see transcript for details
Auto-Submitted: auto-generated (failure)

The original message was received at Fri, 25 Jun 2004 13:42:33 -0700
from localhost.localdomain [127.0.0.1]

   ----- The following addresses had permanent fatal errors -----
<hpsg at lists.stanford.edu>
    (reason: 550 5.1.1 <hpsg at lists.stanford.edu>... User unknown)

   ----- Transcript of session follows -----
... while talking to lists.stanford.edu.:
>>> DATA
Dear HPSG list,

I'm quite pleased at the discussion that has ensued
from my original note.  I wanted to point out that there
are at least two separate questions that we are discussing,
and while they are interrelated, I'd like to make sure
that they are also understood as distinct.

1) What can we do to promote interest in HPSG
more generally?

2) What can we do to promote interest in the
HPSG conference?

In answer to (1), there has been discussion of trying
to present more HPSG work in venues outside the HPSG
conference, including perhaps ending the HPSG conference
as an independent event and discussion of making HPSG
as a theory more interesting to researchers of other
theoretical stripes by developing, researching, and promoting
more theoretical claims.  In response to this latter
thread, I'd like to point out that we collectively are
HPSG, so if you want that, go to it!

Another colleague who wishes to remain anonymous (but
whom I invite to stand up and take credit), sends along
the following list of opportunities for adding "restrictiveness"
at the theory level:

>The real issues about restricting formalisms should be the ones that
>really pertain to making linguistic generalizations.  For instance,
>nobody's proposed anywhere to work on a theory of what kinds of
>features we should have in an HPSG grammar, that is, what kinds of
>attribute/value pairs are motivated for linguistic reasons on general
>principles and why.  That would be linguistically interesting because
>it could restrict the formalism in a way that makes real actual
>predictions about languages.  Or feature geometry.  Why has nobody
>worked on this as a theoretical issue?  People propose whatever
>geometry they like and there's a sort of gentlemen's agreement that
>certain geometries are used, and people argue about sign vs. synsem
>on ARG-ST or whatever, but why not make a theory of it?  Think about
>locality: it's a really, really nice property of HPSG that
>cancellation off of subcat lists in a synsem-based subcat structure
>gives you locality for free, as much as you can believe in locality.
>But it doesn't come off like a neat theoretical prediction to people
>outside of HPSG because nothing forces you to do that.  It's not
>encoded anywhere, and thus isn't so much a prediction as just a way
>things happen to get done.  And finally, type hierarchies: there is a
>fantastic amount of potential here for work in typology if we
>understood and had a linguistically grounded theory of what type
>hierarchies should look like.  As it stands nearly every analysis of
>any phenomenon in any language brings with it these huge, ludicrous
>type hierarchies (and extra features) that are useless for anything
>other than getting the data right.  Big, huge hierarchies proposing
>50+ brand new types for one damn morphological feature of one
>language.  Suppose we had a theory of type hierarchies that says that
>types in a type hierarchy must correspond strongly to a point of
>typological variation.  What predictions might that make, and how
>might it shape HPSG work?  Few people are working on these issues.

I'd like to add to these ideas that I don't think they have
to be "encoded" anywhere in the formalism, especially not in
such a way that one is "forced" to follow them.  Rather, they
can be put forward as hypotheses, which research on various
languages and phenomena can address by noting whether or not
it is consistent with the hypotheses, and in particular whether
it constitutes a counterexample to them (as opposed to just
being developed without regard to them).

In regard to question (2), I think we should perhaps consider
some what we consider the purpose of the HPSG conference to be.
I don't think that promoting interest in HPSG is central to
the purpose of the conference.  Rather, the conference (qua
workshop, per Ivan) serves as a point of contact for members
of the rather dispersed HPSG community.  We work together because
we are working on something (a theory of language) which is too
big for any one person to achieve on their own.  Having a forum
to gather, meet, and exchange ideas can be helpful in that
collaboration.  If, however, the conference is poorly attended
because people are involved in too may overlapping collaborations
and have to pick and choose which meetings to attend, then
something like Dick's proposed "Federated _____ Conference"
might be the way to go.  I have to say that I also like the idea
of topic-focused workshops/conferences, and that I don't feel
like I know the answer to the right balance of those and
framework-focused workshop/conferences.

Finally, since I don't see any harm in making this message
a bit longer at this point, I'd like to respond to another
anonymous contribution:

>I'm surprised to hear you say that we need to get more computational
>linguists in on the picture.  If anything, I think (as Tibor seems
>to) that HPSG is crawling with computational linguists and nobody's
>doing any real linguistics.

I'd like to submit that "real linguistics v. computational
linguistics" is a false dichotomy.  Sure, there is plenty of work
going on under the rubric of "natural language processing" that is
only minimally, if at all, informed by linguistic ideas about language
structure.  (And even some of that can be of interest to linguists: If
an unsupervised tagger consistently produces certain categories of
words based on their distribution with respect to other words,
shouldn't it be worth comparing those results to linguists' part of
speech?)  But, there is also plenty of work, e.g., on linguistic
hypothesis testing through grammar engineering which is a)
linguistically informed and b) provides interesting feedback to theory
developers (often the same people as those doing the implementation,
but not always).

There are even connections with typology:  Together with Dan Flickinger
and Jeff Good, I am starting a project to leverage the LinGO Grammar
Matrix (a language independent core grammar, plus eventually modules
for recurring but non-universal systems), the LKB, [incr tsdb()] and
other tools to assist in the documentation of the grammars (syntax
and morphosyntax) of underdocumented languages.  Paired with these
tools will be others for annotating corpora and creating electronic
descriptive (field) grammars.  Along the way, we will be testing
the various hypotheses encoded in the grammar matrix across a wide
variety of languages.  This work seems to me to be both computational
linguistics and "real" linguistics.

Emily