Two experimental studies on human text processing and knowledge representation are reported. They are designed to explore the nature of cognitive processes in working
memory and long term representations associated with the resolution of reference in texts.
Resolution of references occurs when properties of distinct individuals are bound together
in memory. Stenning Shepherd and Levy (1988) propose that binding is achieved by
recruiting existing general knowledge associations based on the semantic structure of texts.
They present models of representation structures employed in a novel Memory for Individual
Task (MIT), and show that these models can explain certain patterns of retrieval error
frequencies. A statistical model of construction of representation processes which account
for a particular pattern of reading times in terms of key aspects of the structure of MIT
texts is also presented.
The reading times results of the first MIT experiment, in which the order of switches
in reference between individuals is unpredictable, is presented together with as extended
construction processing model which capture phenomena of reference changes. The new
models show that unpredictable reference changes cost time as a function of the complexity
of the individual to which reference is switched, without disrupting the modular
account of processing centred on referenced individual reported by Stenning, Shepherd and
Levy (1988). Analysis of recall errors reveal an effect of presentation order, which results
in confusion over identity of individuals' properties, providing a basis for a distinction
between `primary' and `secondary' individual, each requiring different syllabic rehearsal
processing. These working memory processes are incorporated in a model, which reveals
interaction between rehearsal and semantic processes.
The error data is further analysed with respect to logically constraining solutions to
representations of bindings with `direct' and `indirect' structures. Direct systems represent
binding by structural devices referring to individual identities in their representation;
indirect systems represent binding only through quantificational facts. Both direct and
indirect models are developed and the latter one shown to be at least as good a fit to the
data as the former, which suggests that solution to the binding problem is represented in a distributed manner closer to PDP systems.
Much of the theoretical underpinning of the findings of the first study, is dependent
on aspects of the semantic structures which reflect regularities in the temporal order of
descriptions of individuals in MIT texts. The second study investigates the extent to which
such regularities facilitated the sorts of structures constructed in representations of solutions
to the binding problem, and interaction between temporal order of presentation and
working memory processes. Analyses of reading times show that, while order of presentation
of properties has no significant effect on working memory processes, differences in
availability of information about higher level semantic structure does require extra processing.
A statistical model which factors out some of this processing load as due to specific
changes in the semantic structure of MIT texts is presented.
Detailed analyses of recall error data further reveals differences in indirect representational
features which reflect changes in temporal order of individuals' descriptions.
These models show how texts with the same literal meaning obtain significantly different
representations in memory, not because of different contexts which would be expected, but
due to differences in temporal order of the same sorts of descriptions. This suggests that
the theoretical distinction between working memory processes and long term representations
is not as simple as it might appear. These findings also serve to support our methodological
approach to the study of human text processing.