Filler-gap mismatches

[kaplan at parc.xerox.com]

Hi Carl,

You wrote:

"LFG and HPSG agree in not allowing syntactic selection for dominance
or linear order. The only c-structure property HPSG allows selection for
is category. So one move LFG could make toward HPSG without giving
up its basic architecture is to move category out of c-structure and
into f-structure. You'd still have c-structure as a representation of
dominance and linear order. That would solve some problems, from
my point of view. Is there another reason to keep category labels
on c-structure nodes, aside from the fact that pEhrase markers in
ASPECTS OF SYNTAX looked that way?"

(Another somewhat lengthy reply....)

Actually, it isn't so clear to me that as a matter of principle we want
to exclude lexical selection for linear order.  I think there are cases
in some languages where word order is generally free but particular
predicates might impose more specific restrictions.  Again, in LFG this
can be done without copying linear order into the f-structure, by means
of the f-precedence relation that I mentioned in my previous message,
and again, just as for category, the ordering restrictions are mediated
by the grammatical function assignments--they can't be stated directly
in c-structure terms.  Annie and  I made use of f-precedence in our
account of some of the odd-ball Germanic data (e.g. Swiss German), in
Zaenen & Kaplan (1995) (which also includes a discussion of some of the
functional category issues).  Offhand I don't remember the actual
constraints there, but f-precedence enables you to say things like (^
OBJ) < (^ SUBJ)--and it will impose the ordering restriction among the
nodes that map to those functions even if those nodes are not sisters of
each other.  This may give some of the effect that HPSG would get from
Reape's (or others) domain-union proposal.

So I think it is an empirical matter at this point, not a formal matter
or a matter of principle, whether or not allowing lexical restrictions
on functional precedence will enable a nice separation of ordering
generalizations as stated in general c-structure rules from further
idiosyncratic ordering constraints associated with particular lexical
items.

A couple of comments on the possibility of moving the category out of
the c-structure and into the f-structure.  On a historical note, this
was the way I set things up  in 1975 when I first proposed using
hierarchical attribute value structures instead of trees for underlying
syntactic representations.  But this arrangement caused all the problems
that I alluded to in my last message--it meant that I couldn't simply
conflate the f-structures of nodes along a head chain if they had
different categories.  Instead, I had to keep separate structures
associated with each node (in present parlance, phi would be one-one)
and then define special mechanisms for lifting just the features I
wanted.  By the time I understood enough to define the LFG formalism (in
1978) I had given up on this approach.  The major classification of
properties was into c-structure and f-structure, and ALL f-structure
properties would propagate automatically on a ^=! (up=down) head chain
and even across different nodes representing what we now call co-heads.
So eliminating categories from f-structure seemed like a very good move,
and it has more or less stood the test of time (at least in the LFG
community).

There is another question lurking in your message, which is why bother
having c-structure categories at all?  From a formal point of view, we
know that we could eliminate all c-structure category distinctions
(e.g., a context-free system with a single rule of category X, with a
lots of alternative expansions carrying different grammatical function
assignments and testing different f-structure features).  Indeed, we can
go well beyond that from a formal point of view:  You may recall that
Kelly Roach proved in the early 80's that every LFG language can be
specified by a normal-form grammar that has a single left-branching
phrase structure rule of the form X --> word  (X), essentially a
finite-state c-structure language.  All the constraints that you might
expect to be stated by more specific rules involving order, category,
and dominance can be encoded in a more elaborated set of functional
features.  In fact, given an arbitrary LFG grammar a corresponding
normal-form grammar can be produced automatically that defines the same
set of strings and also assigns feature structures with what was the
original f-structure stashed under a special new attribute.

In a certain sense, such normal-form LFG grammars and structures may
resonate more with HPSG than conventional LFG systems, since essentially
the attribute-value matrix is the only interesting structure and all
constaints are stated in terms of its properties.

This is all by way of saying that there is no formal reason to keep
category labels on c-structure nodes.  But just because it is possible
to move information from c-structure to f-structure doesn't mean it's a
good idea.  It depends on how much explanatory power and descriptive
control comes from maintaining more articulation in the c-structure.
The LFG wager is that there is considerable mileage to be gotten by
making use of a set of named, reusable phrase-structure patterns,
perhaps with universal characteristics, and to think of those names as
available only indirectly for lexical selection.  There might be
difficult or fuzzy cases and perhaps a little patching will have to be
done in various places, but the bet is that the advantages will far
outweigh the disadvantages.

I wonder whether there are specific problems that you are worried about
other than the empirical questions about the behavior of particular
categories that you and Ivan have raised (and which I have not yet
addressed).

--Ron