Two new TRs

[Dr L S Smith (Staff)]

Two new technical reports are available from the CCCN at the University of
Stirling, Scotland. Unfortunately, they are only available by post. To get
them, email lss at cs.stir.ac.uk with your postal address.

TR CCCN-13:

                                                    ISSN 0968-0640



COMPUTATIONAL THEORIES OF READING ALOUD:          
MULTI-LEVEL NEURAL NET APPROACHES


W A Phillips and I M Hay
Centre for Cognitive and Computational Neuroscience
Departments of Psychology and Computing Science
Stirling University
Stirling FK9 4LA
UK


December 1992




Abstract


Cognitive and neuropsychological studies suggest that there
are at least two distinct direct routes from print to sound in
addition to the route via semantics.  It does not follow that
connectionist approaches are thereby weakened.  The use of
multiple levels of analysis is a general design feature of
neural systems, and may apply within phonic and graphic
domains.  The connectionist net for reading aloud simulated by
Seidenberg and McClelland (1989) has elements of this design
feature even though their emphasis was upon the capabilities
of such systems at any single level of analysis.  Our
simulations show that modifying their system to make more use
of its multi-level potential enhances its performance and
explains some of its weaknesses.  This suggest possible multi-
level connectionist systems.  At the lexical level information
about the whole letter string is processed as a whole; at the
sub-lexical level smaller parts, such as heads and bodies, are
processed separately.  We argue that these two levels do not
operate in basically different ways, and that connectionist
and dual-route approaches are mutually supportive.

___________________________________________________________________________
                                                    
TR CCCN-14:


LEXICALITY AND PRONUNCIATION IN A SIMULATED NEURAL NET




W A Phillips, I M Hay and L S Smith
Centre for Cognitive and Computational Neuroscience
Departments of Psychology and Computing Science
University of Stirling
Stirling FK9 4LA
UK


December 1992






Abstract

Self-supervised compressive neural nets can perform non-linear
multi-level latent structure analysis.  They therefore have
promise for cognitive theory.  We study their use in the
Seidenberg and McClelland (1989) model of reading.  Analysis
shows that self-supervised compression in their model can make
only a limited contribution to lexical decision, and
simulation shows that it interferes with the associative
mapping into phonology.  Self-supervised compression is
therefore put to no good use in their model.  This does not
weaken the arguments for self-supervised compression, however,
and we suggest possible beneficial uses that merit further
study.


--Leslie Smith, Department of Computing Science/CCCN University
of Stirling, Stirling FK9 4LA Scotland.

--lss at cs.stir.ac.uk