Joe:
#And Rosta wrote:
#> A.
#> I have given the paper a quick read, and my main response is that
#> the DG/PSG dichotomy is a bit misleading. DG is a notational variant
#> of one particular and particularly-restrictive variety of PSG. Not all
#> PSG syntacticians share the same assumptions, and some espouse
#> something tantamount to DG while others do not. At the same time,
#> some models that have gone under the name of 'Dependency'
#> (e.g. -- from my limited knowledge of different DG models -- Dependency
#> Phonology and Shaumyan's work) are actually notational variants of
#> less restrictive varieties of PSG.
#
#If I recall correctly, in the CG manuscript, Fillmore and Kay say
#somethin like they use PSG, because the facts they want to explain
#can be stated more easily that way. I think that what they're trying
#to say PSG can encompass DG, but not vice versa, the assumption being
#that a PSG can have dependencies but a DG cannot have constituent
#structure (in DG constituents merely fall out of dependency chunks).

So the substantive differences are where the constituent structure is 
such that it doesn't fall out of dependency chunks.

#Strictly speaking, there is a substantive difference that springs
#immediately to mind: bracketing paradoxes. Take for example:
#
#1. I gave Mary some candy and Joan some flowers.
#
#On any reasonable analysis of constituent structure, [Joan some
#flowers] is not a legal constituent. 

Well, IMO it is a constituent, but that's another story. Certainly the
consensus is that it isn't a constituent.

#A DG would have no problems
#with this, but PSG would. In fact, CG would of course have to
#propose a special construction to deal with this.

I really don't see how a DG fares any better. The essence of the
problem is that the coordinate and noncoordinate structures
require two different nonconverging structures. WG does this
by assigning constituent structure to coordinate structure and
dependency structure to noncoordinate structure, but a model
could equally well use dependency for both or constituency
for both.

#> The real substantive parameters of differences are these:
#>
#> 1. Is total or partial exocentricity allowed? That is, are the features of a
#> nonterminal (or maximal) node anything other than a complete copy of
#> the features of its lexical head? DG says No, and it is this that most
#> differentiates so-called Dependency models from PSG models.
#
#This can't be true of coordination. For example:
#
#2. Ray and George were discussing the matter very politely.
#
#Ray has to have the feature [num sg] and so does George. Yet the
#coordinate structure as a whole has to be marked [num pl]. Perhaps
#a "pure" DG account can't explain this, which is why WG adopts
#something more akin to constituency with coordination.

What I say under (1) is correct, but you're right that DG then has
a problem with (2). My solution would be to take the head of NP
coordination to be the conjunction (so _and_ is subject of _were_).
WG's use of constituency doesn't help here, because the subject
of _were_ is both _Ray_ and _George_ in a WG analysis.

#> 2. Can a head node be head of more than one branching node (as
#> happens with multiple bar levels and with adjunction)? Pure DG says
#> No. But a pure DG model that allows labelled branches can translate
#> a model that allows a head node to be head of more than one
#> branching node (e.g. by using such branch labels as 'Adjunct' or
#> 'Specifier'), so the substantive difference comes down to the extent
#> to which the extra nodes yield the correct word orders and interpretations.
#
#Could you give an example? If you're thinking of something like:
#
#3. NP
#   |
#   N'
#   |
#   N
#   |
#  she
#
#where you have two "empty projections," CG doesn't allow this
#(pronouns in CG are [max +, lex +]. In fact, except in cases of
#conversion (i.e., count-->mass), CG says that every mother must
#have at least two daughters. In this sense, CG hopes that
#constituent structure will be highly consistent with "surface
#structure."

No; I mean where mothers have at least two daughters, but
where a nonterminal node's lexical head is not necessarily
its daughter, i.e. where X is a projection of Y but not a
mother of Y. Examples are specifiers (XP > spec + X') and 
adjunction (XP > XP + YP). As I said, DG could simulate this 
by treating 'spec' and 'adjunct' as dependency labels, but
the PSG makes predictions/claims about word order and
semantic phrasing that the DG doesn't.

--And.