Thesis available: Syntactic Innovation--a Connectionist Model

[Whitney Tabor]

The following dissertation is available by ftp.  There are two
alternative formats: tabor.thesis.2up.ps.Z (120 sheets of paper),
tabor.thesis.ps.Z (230 sheets of paper).

Ftp host:  csli.stanford.edu
Ftp filename:  pub/Preprints/tabor.thesis.2up.ps.Z
Ftp filename:  pub/Preprints/tabor.thesis.ps.Z


           Syntactic Innovation:  a Connectionist Model

                          Whitney Tabor*
                    Department of Linguistics
                       Stanford University
                              1994

This thesis uses the continuous representation space of a
connectionist network to model syntactic innovation in natural
language.
  
Current theories of grammar, following Chomsky 1957, generally take
quantitative properties of language use (e.g., context-dependent
word-frequency information) to be irrelevant to the determination of
grammatical structure.  Such theories are not very useful for making
constrained predictions about historical grammar change for the
changes they can model are unrealistically abrupt and there seems to
be no domain-independent structuring of the representation space that
reveals which changes are probable and which are not.  But recent work
in the field of grammaticalization (e.g., Traugott and Heine 1991,
Hopper and Traugott 1993) indicates strong, general constraints on
grammar change.  Moreover, these constraints often have quantitative
correlates: changes in the categorical status of a word or morpheme
are often preceded by significant shifts in its relative frequency
distribution.

I show how a recurrent connectionist network trained on word
prediction in the manner suggested by Elman 1990 and 1991 is sensitive
to the information provided by these quantitative shifts and thus can
be used to predict correlations between the frequency changes and
categorical changes.  The correctness of these predictions is borne
out by case studies of degree modifier <sort/kind of> and future <be
going to>.  Two additional empirical phenomena support the model:
linked frequency changes in grammatically related constructions and
emergence of hybrid structures during periods of transition.  The
model permits simplification of the theory of language change by
replacing "reanalysis" and "analogical extension" with a single type
of change-event.  It also permits simplification of the theory of
grammar by allowing the same interpolative mechanism to handle both
normal productive syntax and the hitherto problematic hybrid cases.


*Current address:

Whitney Tabor
Department of Psychology
University of Rochester
Meliora Hall---River Campus
Rochester, NY  14627