theory and formalism in LFG: too much inertia? (even longer message!)

[Paul King]

Dear all

Let me preface my (long) posting with two apologies and a caveat.
First the apologies.  I intended to reply to Prof. Bresnan's posting
much sooner, but other matters unavoidably intervened.  I can only
hope that the topic has not already cooled.  Moreover, I seem to be
going down with a cold or some such, and feel that my intellectual
faculties are not at their best.  Now, the caveat.  I work primarily
within the HPSG paradigm, and as such my opinions concerning the LFG
paradigm are worth significantly less than the opinions of those that
earn their bread and butter by it.

I am drawn to reluctantly raise my head above the paradigmatic parapet
for two distinct reasons.  Firstly, the posting of Prof. Bresnan risks
perpetuating a misguided conflation of notions.  Secondly, I fear that
what Prof. Bresnan writes has at least the potential to destroy some
of those aspects of LFG that I consider most admirable.  Both are, in
a very real sense, orthogonal to Prof. Bresnan's concerns.  However, I
have seen sufficient damage done within the HPSG community by leaving
the potentially misleading unchecked that I cannot conscionably keep
silent and risk similar damage occurring within the LFG community.

My first point concerns the role of formalism and theory in
linguistics.  I view linguistics -- at least the kind of linguistics
performed by LFGians and HPSGians -- as a deductive science, in which
scientists (linguists in this case) propose explicit hypotheses
concerning a certain aspect of the universe (natural language in this
case) which are then tested by (i) deducing from the hypotheses
explicit conclusions and (ii) subjecting a challenging sample of those
conclusions with empirical content to experimental test.

Within this conception of science in general and linguistics in
particular, a theory is a collection of hypotheses, and a formalism is
a mechanism whereby that theory can be explicitly and unambiguously
expressed.  For example, Maxwell's equations of electromagnetism
constitute a theory.  The formalism in which that theory is expressed
is vector analysis.  Note, vector analysis does not itself constitute
a theory of electromagnetism.  Indeed, I have seen theories of
economics expressed using vector analysis.  Rather, vector analysis is
merely a vehicle with which Maxwell's theory can be expressed.  If
Maxwell's equations seem too abstruse or remote, consider HPSG
instead.  A HPSG theory is a set of principles, such as the head
feature principle.  One formalism in which that theory can be
expressed (there are alternatives) is an object constraint language.
Note, the constraint language does not itself constitute a theory of
natural language.  Rather, the constraint language is merely a means
whereby the principles that *do* constitute the grammar/theory can be
faithfully and unambiguously expressed.

Many linguists fail to clearly distinguish formalism and theory, and
it subsequently costs me a great deal of teaching effort every
semester to wean students with some linguistic experience from a fear
of powerful formalisms.  That Chomsky conflated formalism and theory
(perhaps for good reason, perhaps not) should not blind linguists to
the facts that this conflation is neither obligatory nor typical.  To
extend my above examples, vector analysis allows for more than the
three spatial and one temporal dimension in which Maxwell's equations
are couched, yet no-one would accuse Maxwell's theory of being too
weak simply because the formalism is too strong.  Some constraint
languages allow for the determination of any recursive set of strings,
yet no-one should accuse any particular HPSG grammar of being too weak
simply because the formalism is too strong.  All physical, biological
and economic theories I am aware of use formalisms that have the
potential to express much, much more than the individual theories
actually express.  It is the role of the theoretical scientist, not
the formalism, to carefully enunciate their insights regarding the
universe.  To claim that a formalism must enable a theoretical
scientist to say what they want to say *and no more* is as absurd as
to claim that a soccer journalist should not use the English language
to report on tonight's game between Manchester United and Olympique
Marseilles because the English language enables discussion of
philosophy.[footnote 1]

Unfortunately, I find Prof. Bresnan's posting a little ambivalent as
regards the relationship between formalism and theory.  On the one
hand, I believe that Prof. Bresnan is rightly distinguishing formalism
and theory when she writes that a formal architecture "will be useful
if you can use it to model a very wide range of hypotheses" and chides
those linguists who "try to have a very restrictive theory, and build
the restrictions into their formal architecture."  On the other, I
believe she is wrongly conflating formalism and theory when she writes
that "In this sense, a powerful formalism is a weak theory."  The
claims that "LFG has separated its very powerful formalism (...) from
the various substantive theories" and "[t]he latter have generally
been treated as metatheories, and have seldom been incorporated into
the formal architecture" further muddies the waters.  This could be
read as endorsing the claim that a formalism is merely a mechanism for
expressing theories (and is possibly capable of expressing much more).
But the curious use of the term "metatheory" and the notion of
"incorporating" metatheory into formal architecture could equally well
indicate that it is the metatheory that establishes the terms of
reference for a theory and the formal architecture that exploits those
terms of reference to make explicit claims about language.

The final paragraph of Prof. Bresnan's posting -- with its ringing
call to examine and extend if necessary the LFG formalism --
encourages me to believe that the earlier conflations of formalism and
theory are slips of the keyboard or misunderstandings on my part.
Certainly, I can only concur with Prof. Bresnan in her call.  However,
the call leads to my second reason for risking the ire of the LFG
community in general and Prof. Bresnan in particular.  And this reason
is a genuine cri du couer.

As part of a new research project, it is my duty to assemble a list of
significant readings concerning the philosophical and mathematical
foundations of both LFG and HPSG.  It has been an honest (if time
consuming) pleasure to read again texts I have not studied for nearly
a decade, and to read anew texts I had never seen before.  Among the
many impressions that have formed in my mind during this intense
period of scholarship is the extent to which the development of early
LFG was driven by an explicitly stated principle.  Onward from
Prof. Bresnan's work in Lexical Generative Grammar and Prof. Kaplan's
work in Augmented Transition Networks, the leit motif of LFG has been
to fashion a grammatical paradigm that can accommodate psychological
realism.  LFG is in sharp contradistinction to HPSG here.  I must
confess I find HPSG to have little if any underlying principles, other
than those it inherited from GPSG.  It might not be a coincidence that
whereas HPSG has spawned a panoply of formalisms -- some compatible
with others, some not -- the LFG formalism has evolved more gradually,
more incrementally, as suggested modifications to its underlying
formalism have been found to accord with its underlying principle.

I am therefore more than a little sad and disturbed at what I see as a
possible call to tacit abandonment of this guiding principle.  I
simply find it hard to reconcile some of the formal modifications
suggested in Prof. Bresnan's posting and elsewhere in the LFG
literature with LFG's underlying principle.  Perhaps the synthesis of
these modifications and psychological realism is transparent and some
shortcoming of mine prevents me from seeing it.  But I have been
working on foundational issues of linguistic paradigms for so long now
that I (perhaps arrogantly and erroneously) find this hard to credit.
Rather, I suspect that some of the proposed modifications are simply
incompatible with psychological realism, or their compatibility is
sufficiently opaque that it warrants demonstration.  Or perhaps the
principle of psychological realism *should* be abandoned.  However, if
so, then LFGians should at least be clear and honest about what is
being abandoned, what is replacing it, and what the consequences might
be.

This is not a call to stagnation, but an injunction against change for
change's sake.  If we had adopted all of the ideas that have been
first lauded and then left by the pundits of intellectual fashion over
the past decade and a half, we would currently be peddling situated,
stochastic, chaotic and resource-bound catastrophic unification
grammar (with non-well-founded sets).  Each of these notions is
worthwhile in itself, but it does not follow that each is appropriate
for LFG.  Rather, LFG should do as it has always done: enunciate its
underlying principles and modify them as needed, fashion a formalism
appropriate to those principles, and then do good linguistics using
the formalism.

Let me express my gratitude to those of you who have read this far.
------------------------------
[footnote 1]	An interesting side-issue would be to ask why this
		conflation is so prevalent in linguistics.  I suspect
		the reason is that linguists do not always clearly
		distinguish between a grammar as a scientific theory
		of natural language and a grammar as a natural
		language processing engine.  Indeed, depending upon
		your particular conception of what a natural language
		is, it might be correct *not* to distinguish grammars
		in this way.  However, just as with the conflation of
		theory and formalism, the conflation of theory and
		process is neither obligatory nor typical.  A theory
		of arithmetic need not faithfully capture the process
		humans use to calculate their tax-returns.  However,
		this is a side-issue in a very long posting that
		already departs from the main thrust of
		Prof. Bresnan's post, so I will not pursue this issue
		further here.

Tschuess
Paul
______________________________________________________________________
Paul John King __ The CLaRK Programme, Seminar fuer Sprachwissenschaft
Mail _________________ Kl. Wilhelmstr. 113, D-72074 Tuebingen, Germany
Email ________________________________ king at sfs.nphil.uni-tuebingen.de
WWW ______________________ http://www.sfs.nphil.uni-tuebingen.de/~king
Telephone _______ Vox (0/+49) 7071 2978490 and Fax (0/+49) 7071 550520