theory and formalism in LFG: too much inertia? (long msg)

[Yehuda N. Falk]

From: Mark Johnson <mj at cs.brown.edu>
Subject: Re: theory and formalism in LFG: too much inertia? (long msg)
To: msyfalk at mscc.huji.ac.il
Date: Fri, 24 Sep 1999 11:48:52 -0400 (EDT)
Cc: LFG at LINGUIST.LDC.UPENN.EDU
X-Mailer: ELM [version 2.4ME+ PL37 (25)]
MIME-Version: 1.0
Content-Type: text/plain; charset=US-ASCII
Content-Transfer-Encoding: 7bit


> The fact is that the current LFG formalism does not always allow the
> expression of markedness- (or whatever-) based generalizations. This is an
> indication that the formalism does need to be enriched. But there is also a
> side to this that Joan didn't mention: a formalism can also allow us to put
> constraints on what is a possible substantive linguistic constraint.

As someone who's worked all my professional life on grammar
formalisms, I now think that they shouldn't be viewed as the central
part of any linguistic theory.  Rather, the most important, central
claims of a linguistic theory are things like: the nature and role of
grammatical relations, the role of the lexicon, the kinds of
linguistic relationships we posit, etc.  In order to explicate these
we devise a formalism, and indeed, I think we should judge a formalism
by how well it permits us to define and explain the pre-formal central
claims of the theory.

If this is correct, it should be possible to have several different
formalizations of the same linguistic theory.  For example, I proposed
an alternative formalization of LFG based on linear logic in a series
of recent papers: among other changes this formalization gives up the
requirement that features appear in attribute-value pairs, e.g.,
CASE=NOM (which I assume was never a central _linguistic_ claim of
LFG), and shows how to make sense of unattributed features such as NOM
standing alone.  However, I claim that this is still a version of LFG
because it formalizes key linguistic notions such as grammatical
relations, c-structure and f-structure, etc.  Even more, by saying
that my formalization is a formalization of LFG, I am implicitly
saying that you can criticize my formalization by pointing out a
central idea of LFG that my formalization doesn't provide (e.g., one
thing my formalization doesn't yet support is the LFG view of
coordination as involving a grouping of the coordinated phrases over
which the grammatical relationships distribute).

> It is also true that the place of OT, probability, etc. in LFG (or any
> formal linguistic theory) is still up for grabs. We do not know the full
> extent of these elements or how they interact with the rule and
> representational systems that we all know and love.

I think that the problem is far harder (and far more important) than
this!  It's actually fairly easy to define probability distributions
over LFG structures (see my Association of Computational Linguistics
1999 conference paper for a description of one way to do this in a
computationally tractable way), and there are surely many other ways
too.  (Ron Kaplan has been working on a version of Bod's DOP, for
example).

But just because it can be done formally doesn't mean that it makes
sense linguistically.  Of course there's an interplay here -- the
mathematics can suggest ways in which notions like OT or probability
might interact with linguistic phenomena, but if the theory we come
up with isn't correct linguistically, I think its formal qualities
are rather moot.

Well, that's my 2c worth!

Best

Mark

PS.  My papers are available off my Web site http://cog.brown.edu/~mj
I love to have comments, positive or negative.