[Prev][Next][Index]

XTAG meeting on Feb 10th, 10:30am


Hi all,

Seth Kulick will be giving a pratice (job) talk this thursday during the
Xtag meeting. The audience will probably be larger than the usual since
everyone (well... everyone within the assumed domain) has been invited in
a separate mail.

VENUE: large conference room.

His abstract follows:

---------------------------------------------------------------

	Constraining Non-Local Dependencies in Tree Adjoining Grammar:
		Computational and Linguistic Perspectives

				Seth Kulick

The linguistic constructions sometimes loosely grouped under such terms as
"reduced constructions" or "restructuring" have long been of interest for
syntactic theory.  Cases of reduced constructions such as clitic climbing
and the long reflexive passive in Romance, or long distance scrambling in
German, appear to occupy an intermediate position between local and
long-distance movement, thus posing a challenge for the proper statement
of locality constraints.

In this talk I examine reduced constructions from the perspective of Tree
Adjoining Grammar (TAG).  TAG is not a linguistic theory; rather, it is a
mathematical formalism which can be used to encode various linguistic
theories.  The distinguishing property of TAG is the factoring out of
recursion from the statement of the grammar, thus forcing the substantive
theory of syntax to be stated over bounded, local domains called
"elementary trees".  Various syntactic locality constraints fall out from
the restricted nature of a TAG derivation.

Previous work on long distance scrambling and TAG has resulted in
extensions of TAG which, I argue, suffer from an abandonment of this
restrictiveness.  Previous work on clitic climbing has resulted in systems
which maintain the restrictiveness but which cannot handle certain cases
of clitic climbing.  Furthermore, given the similarities between clitic
climbing and long distance scrambling, it is undesirable to have two
different extensions of TAG to accomplish similar goals.

In this work I propose a revised system to handle these problematic cases
without losing the desirable properties of TAG.  There are two main
aspects to the solution.  First, a closer examination of some of the core
cases previously applied to TAG (wh-movement, subject-to-subject raising)
reveals that the problems caused by reduced constructions can be seen as
just another case of an underlying structural problem already existing in
TAG, leading to a natural generalization of the adjoining operation.
Second, unlike the previous work on scrambling and TAG, a strong notion of
locality is retained in the derivational structure.

The resulting system, "Segmented Adjoining Grammar" (SAG), is of interest
for several reasons:

1. As a natural generalization of TAG, the locality constraints on clitic
climbing and long distance scrambling fall out of the structure of the
derivation, in the same way as such cases as "super-raising" fall out of a
TAG grammar.  The problem of the statement of locality constraints on
these various cases of movement is thus greatly simplified as compared to
other systems, since there is no need for any such statement of
constraints.

2. It correctly predicts some cross-linguistic differences between clitic
climbing and long distance scrambling. Despite their similarities in other
ways, there is a constraint against what Aissen & Perlmutter 1976 called
"intersecting clitic climbing" while, as I show, long scrambling allows an
analogous type of movement. This follows under my analysis entirely from
the representation of clitics vs. scrambled NPs in an elementary tree,
without the need for any stipulations about how clitics vs. NPs have
different interclausal movement constraints.

3. It leads to some interesting questions about the competence/performance
distinction.  Locality is retained in SAG by sacrificing coverage of some
theoretically possible cases of long scrambling that never actually occur,
due to their excessive complexity, such as having many NPs simultaneously
long scrambling in one sentence.  While the exact extent of this loss of
coverage is still under study, such cases are usually ruled out by
"performance" considerations.  This work shows that such cases can instead
fall outside the "competence" grammar itself.

4. Approaches to reduced constructions in other frameworks usually use the
same basic mechanism for the various aspects of these constructions in one
language.  For example, a clause union or argument composition analysis
will usually handle both the long reflexive passive and clitic climbing.  
This is not true for a formalism in the TAG family, whether the one I
propose here or another one.  Instead, sharply different analysis must be
given for the different constructions, thus making certain predictions
about the ranges of verbs that allow the different cases of reduced
constructions.


---------------------------------------------------------------

- Rashmi