Abstract
Research in Sentence Processing is concerned with discovering the mechanism by
which linguistic utterances are mapped onto meaningful representations within the
human mind. Models of the Human Sentence Processing Mechanism (HSPM) can
be divided into those in which such mapping is performed by a number of limited
modular processes and those in which there is a single interactive process. A further,
and increasingly important, distinction is between models which rely on innate
preferences to guide decision processes and those which make use of experiencebased
statistics.
In this context, the aims of the current thesis are two-fold:
• To argue that the correct architecture of the HSPM is both modular and
statistical - the Modular Statistical Hypothesis (MSH).
• To propose and provide empirical support for a position in which human
lexical category disambiguation occurs within a modular process, distinct
from syntactic parsing and guided by a statistical decision process.
Arguments are given for why a modular statistical architecture should be preferred
on both methodological and rational grounds. We then turn to the (often ignored)
problem of lexical category disambiguation and propose the existence of a presyntactic
Statistical Lexical Category Module (SLCM). A number of variants of the
SLCM are introduced. By empirically investigating this particular architecture we
also hope to provide support for the more general hypothesis - the MSH.
The SLCM has some interesting behavioural properties; the remainder of the thesis
empirically investigates whether these behaviours are observable in human sentence
processing. We first consider whether the results of existing studies might be
attributable to SLCM behaviour. Such evaluation provides support for an HSPM
architecture that includes this SLCM and allows us to determine which SLCM
variant is empirically most plausible. Predictions are made, using this variant, to
determine SLCM behaviour in the face of novel utterances; these predictions are then
tested using a self-paced reading paradigm. The results of this experimentation fully
support the inclusion of the SLCM in a model of the HSPM and are not compatible
with other existing models.
As the SLCM is a modular and statistical process, empirical evidence for the SLCM
also directly supports an HSPM architecture which is modular and statistical. We
therefore conclude that our results strongly support both the SLCM and the MSH.
However, more work is needed, both to produce further evidence and to define the
model further.