Filler-gap mismatches

[Carl Pollard]

Dear Yehuda Falk,

With all due respect, I disagree with the following statement:

>
I think the crucial difference between the theories is the one Ron has
identified: modularity. LFG is crucially based on the idea that different
dimensions of linguistic structure (constituency, function, phonology,
argumenthood, semantics, information structure, etc.) are represented at
distinct levels with there own formal properties, primitives, principles of
organization, etc.,
>>

The same is true of HPSG: constituency is represented by one set of
features (the `daughters' features HEAD-DTR, SUBJ-DTR, etc.), function
by another set (the subcategorized valence features SUBJ, COMPS, SPECIFIER,
and the head-selecting valwnce feature MOD), phonology by the
PHONOLOGY feature, argumenthood by the ARG-ST feature, semantics by the
CONTENT feature, and within each set of features, there are distinct
types (species) of values, distinct constraints that apply, etc.
The fact that all these things are stuck together in a common large
strcuture is no more significant than if we had stuck them all together
into an ordered n-tuple.

>
and that these parallel levels are related to each
other by correspondence relations like the phi function mapping c-structure
to f-structure.
>>

Likewise, the different levels in HPSG are related by interface
constraints.

In my view the decision to model everything in HPSG by feature
structures is essentially a stylistic choice. It all could have been
done using the first-order language of set theory instead of feature
logic, or it could have been done more like LFG, using trees for the
daughters-type information, making phonology-syntax-semantics an
ordered triple instead of a feature structure, etc. None of this is
very significant.  At the moment I am developing a framework where the
universe for modelling linguistic entities is neither the world of
feature structures nor the world of sets, but rather certain kinds of
categorical models of higher-order logic. One reason for doing this is
because I think it makes the (well-concealed) points convergence among
HPSG, LFG, and type-logical grammar more transparent (e.g. the
Levine-Hukari-Calcagno theory of feature indeterminacy, the
Bayer-Johnson one, and the Dalrymple-Kaplan one are all seen to be
essentially the same theory at a suitable level of abstraction).

>
So the basic conceptual reason for not including category information in
f-structure is that categories are information about the distribution of
elements of overt expression (which is what c-structure models),
>>

Aha, this is a point of convergence between LFG and categorial
grammar.  This is a genuine difference between those two frameworks
and HPSG.  In HPSG, categories (once valence information is removed
from them) are essentially parts of speech, but these have more to do
with paradigmatic than syntagmatic relations; the latter are
determined by valence, not part of speech, and order is determined by
interface constraints between syntax and phonology.

>
not functional information.
>>

I believe HPSG practicioners tend to be mystified by the LFG notion
`functional', since the attributes that appear in an f-structure
seem to be a mix of syntax-semantics interface (PRED), valence
(SUBJ, OBJ, etc.) and morphosyntax (CASE, PFORM, etc.)

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?

>

If the notion of modularity that LFG is based on is right, this
distinction will bear fruit. So the fact that choosing the category of
an XCOMP (like "become" does) seems to be a more complex option is
accounted for by the fact that it requires reference to corresponding
elements on a different representation.
>>

>From the point of view of HPSG practicioners, BECOME is typical, not
exceptional! It is not clear to me that making reference to a
different representation rather than another part of the same
representation (being `a more complex option') has any EMPIRICAL
significance. It seems a little bit like having a logical theory and
writing some of the axioms in Helvetica and some of them in Times
Roman.

>
Cross-pollination of ideas between theories
is a good thing, and there certainly are many similarities between LFG and
HPSG as constraint-based, lexicalist, unificationist theories. But LFG and
HPSG are not notational variants of each other. Ultimately, empirical
evidence will hopefully point the way to which theory is closer to the
"truth"
>>

I hope not and think not. What I hope and think will happen is that
further evidence will resolve conflicts between the two frameworks and
cause them to both evolve into things that are more similar to each
other. I cannot buy the view that there are two (or more) essentially
static total packages and we are waiting to see which total package is
best.

>
The best line to take is for proponents of
each theoretical framework to develop the best analyses they can, and then
we can see where they go.
>>

Again, I respectfully disagree. I have less than zero interest in
holding HPSG fixed and trying to develop the best possible HPSG
analyses, and boundless interest in trying to develop successor
frameworks which incorporate the best features of frameworks like
type-logical grammar, LFG, and HPSG. When I teach the HPSG intro.
course at Ohio State, I always make clear that learning to use HPSG
to analyze data is a secondary goal, and that what I am really trying
to teach them to do is to make HPSG and other contemporary frameworks
obsolete, by laying the groundwork for developing the frameworks
of the future.

>
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).

Carl