Filler-gap mismatches

[kaplan at parc.xerox.com]

Hi again,

The idea of putting just the part-of-speech in the f-structure and not
the bar-level does remove conflicts that would otherwise appear up a
single head chain, but it still eliminates some of the descriptive and
explanatory advantages of the current set up.  For example, we can
account for constructions such as gerunds that have a mixture of
properties that are normally associated with different parts of speech
by saying that nodes of the different part-of-speech (ignoring
bar-level) types map to the same f-structure, in effect, that the
c-structure nodes are co-heads of the functional units.  (I think Joan
and others have written about this).

A significant point about descriptions through inverse correspondences
is that the induced properties (f-category, f-precedence) are degraded,
less informative versions of the related properties in c-structure.  The
inverse correspondence fuzzes things up a bit, because the forward
correspondence is not one-to-one and not onto.  We claim that the
induced fuzziness is just what you need to naturally characterize
different kinds of phenomena.  Now, you could get the same effect by
saying that a part-of-speech attribute in the f-structure is not filled
by a single category symbol but rather by a set of categories, and then
various mixture properties could be enforced by testing for one or
another element of the set.  But that would be a pure stipulation
independent of anything else in the formal system; doing it not by
copying but by making information accessible looking backwards makes for
a stronger and more natural set of claims.

But even if we were to make the move of copying part-of-speech
categories, I'm not sure that the HEAD and ARG-STRUC combination would
then correspond directly to what we have in f-structure, although I
agree they would be more similar.  (It may be that I don't understand
exactly what you (or the correct theorists) code in those features).
The LFG f-structure has features like tense and case, and also modifiers
like adjuncts, which are not strictly speaking arguments (probably the
opening gambit of another long discussion).  Does ARG-ST also include
adjuncts and other kinds of features?

Another historical note:  again in the mid 70's I also included
arguments in the feature structure under special attributes like ARG1,
ARG2, so that everything would be in a single representation.  But again
that was abandoned, mostly in the interests of modularity.  We wanted to
have a representation for syntactic properties that encoded only the
semantic information that interacted with other syntactic features and
was a sufficient basis for later full-scale semantic interpretation.
The PRED semantic-forms carry the minimal set of properties:  the
semantic arity, the mapping of arguments to grammatical functions, and
the instantiation of individuals.  This is sufficient, we suggested, for
working out a lot of syntactic stuff and for many syntax/semantics
interactions, but without committing to any particular theory of
semantic representation or semantic interpretation (the "glue" semantics
now gives us a theory of interpretation, but even it can still leave
open the question of representation--modularity strikes again).

The argument list in a semantic form may be the thing that corresponds
directly to your ARG-ST--and it is typically represented as an ordered
sequence, which might meet one of Berthold's reservations.  But I don't
think we have the worry about composition of ARG-ST values on the mother
of coordination, the point from the thread a few months ago.  (Indeed, I
don't think most LFG people would understand what the problem was that
that discussion was all about--we were going to have a little study
group at Parc to go over all the messages and see if we could map any of
the issues onto problems in LFG.)

--Ron