Filler-gap mismatches

[Carl Pollard]

Hi Yehuda,

>
I am not equipped to debate the formal similarities and differences between
the LFG and HPSG approaches, so I'll leave that to Ron :) But it seems to
me that representing everything as feature structures (AVM) is a claim
that, while the specific features may be different, all these different
dimensions are essentially the same.
>>

No, it really is not intended to make such a claim. Not anymore than
writing a program in LISP is making a claim that everything is the
same because `they are all lists' or modelling in set theory is
claiming that everything is the same `because they are all sets'. I
now regret the choice of feature structures for modelling linguistic
entities precisely because they seem to invite misunderstandings of this
kind. (I am blaming myself, not you, for this misunderstanding.)

What was nice about modelling different kinds of things with similar
data structures was that all interface constraints could be written in
the same language (say, RSRL).  Now I am inclined to give that
advantage up and use, for example, different logics (though both
higher-order) for syntax vs. semantics, and another logical device (a
kind of labelled deductive system) for the syntax-phonology and
syntax-semantics interface. In this respect, I am taking Ron's advice
(using different descriptive devices for different kinds of things),
though probably in ways that will look too much like categorial
grammar for the taste of either LFG or HPSG folks (and probably
too much like HPSG and LFG for the taste of the Lambek-CG folks).

One advantage of higher-order logic over feature logic as a medium for
syntax is that it provides ways of making things that (arguably) ARE
different LOOK different. For example, one can treat bundles of
morphosyntactic features as named products (mathematically the same
kind of thing that f-structures are), so that the features are
essentially projections; while treating `grammatical argument'
features like HPSG's SUBJ, COMPS, SPR, etc.  or LFG's SUBJ, OBJ,
etc. via exponentials. (The difference is analogous to the difference
between cartesian products and function spaces.)

Another advantage is that the categorical models are abstract in a way
that eliminates inessential differences between different concrete
modelling choices, e.g. whether to say that feature structures are
functions that have features as arguments or that features are
functions that have feature structures as values. This is a point that
was argued at length in a series of papers by Drew Moshier in the
mid-90's on using type theory for linguistics.

>
>What do you take to be the sense of `functional' that unites
>these things? What do (say) SUBJ and (say) PFORM have in common
>that makes them be grouped in the same module?

>
LFG divides f-structure attributes into grammatical functions and features.
Grammatical functions are "functional" in the sense that they specify the
purpose(s) of various elements in the syntax. Some elements in the syntax
function to express the arguments of predicates (SUBJ, OBJ, the OBL family
of functions, etc.), some function as adjuncts, and some to express
relationships between the clause (or elements thereof) and larger syntactic
units (SUBJ) and discourse units (FOCUS, TOPIC).

Features provide the content of the functions (the value in the AVM). The
PRED feature ties the argument-taking properties together and represents
the general meaningfulness of elements in syntax (not their meaning, which
is semantics, but the fact of their meaningfulness, which is relevant for
how things function in the syntax). In a sense, the PRED is the f-structure
"head". Other features, such as PERS, NUM, and GEND, provide other aspects
of the content of the functional elements of a sentence. Features like CASE
(and PCASE, if it really is distinct from CASE) help determine the function
of the element of which they are a part.
>>

But there is nothing in either the LFG or the HPSG *FORMALISMS* that
says that functions are different from features. In that respect (but
not in others) I am in favor of moving toward something that is more
similar to categorial grammar.

>
> >
>Trying to make LFG look more like HPSG (or vice
>versa) serves no purpose.
> >>
>
>The point is not to make one look more like the other but to try to
>resolve differences and converge on a common understanding. In fact,
>HPSG originated as an effort to resolve differences between GPSG, LFG,
>GB, and pre-Lambek categorial grammar (which may be why practicioners
>of all of these have accused HPSG of being a notational variant of
>their framework).
>>

>
I am not sure we really disagree here, although maybe you think we are
farther along than I do. To tie this to the earlier thread on
P&P/GB/MP/transformational syntax/Chomskyan theory (or whatever one
wants to call it), I think it is important for all of us to be aware
of developments taking place in as many theoretical frameworks as we
can.
>>

Can't argue with that!

>
I think it is premature to "resolve differences and converge on a
common understanding", because I think there are too many facts that
are not well enough known for there to be a common understanding of
their import.
>>

What I said was to *TRY* to resolve differences and converge on a
common understanding. It is a goal that will never be achieved because
the one true theory can never exist, but I don't think it is ever
premature to TRY to do this, not any more than I think (pace Chomsky)
that it is `premature' to try to formalize linguistic theory.

>
The filler-gap mismatches which the subject line of this
thread refers to are a case in point. They are old news for LFG, and
not considered a problem.  However, they are apparently (at the very
least) not as well known in HPSG, and I would imagine even less known
in derivational circles. So I don't see that we are yet ready to
converge. In the meantime, the various frameworks *do* make different
claims and different predictions, and it is worth taking the time to
test them out.
>>

I would say it is the THEORIES proposed WITHIN the frameworks that make
claims/predictions, not the frameworks themselves. I think it
would be very hard to write a theory in EITHER framework such that
a theory with the same predictions could not be written in the
other. A theory can be easily refuted by facts, but it is much harder
(if not impossible) to refute a framework!

>
This doesn't mean keeping the theory static, either. The replacement
of the original LFG "constituent control" formalism for filler-gap
constructions with functional uncertainty is a nice example of a
change that has resulted from exploring the predictions of the theory.
>>


That's changing the framework, not the theory (isn't it)? (Actually,
it is not quite clear to me whether LFG makes as clear a distinction
between the two as HPSG does.) In any case, you are right.

Carl