Thesis placed on Neuroprose archive

[Bruce Lambert]

****Please do not forward to other lists ***********
 
The following paper has been placed on the neuroprose archive as
 
 lambert.thesis.ps.Z
 
A CONNECTIONIST MODEL OF MESSAGE DESIGN
Bruce L. Lambert
Department of Speech Communication
University of Illinois at Urbana-Champaign, 1992
 
Complexity in plan-based models of message production has typically been
managed by the systematic exploitation of abstractions.  Most signifi-
cantly, abstract goals replace concrete situations, and act types replace
instantiated linguistic contents.  Consequently, plan-based models
involve three key transformations: (a) the mapping of concrete situations
onto abstract goals, (b) the mapping of abstract goals onto act types,
and (c) the mapping of act types onto linguistic tokens.  Transformations
are made possible by the prior functional indexing of situations, acts
and utterances.  However, such functional indexing schemes are based on
untenable assumptions about language and communication.  For example,
communicative means-ends relations are not fixed, and the assignment of
functional significance to decontextualized linguistic forms contradicts
a substantial body of evidence about the contextual dependency of
language function.
     A connectionist model of message production is proposed to avoid
these difficulties.  The model uses Quickprop, a modified version of the
backpropagation learning algorithm, to learn a mapping from thoughts to
the elements of a phrasal lexicon for a hypothetical broken date
situation.  Goals were represented as distributed patterns of context-
specific thoughts.  Messages were represented as distributed patterns of
phrases.  Three studies were conducted to develop and test the model.
The first study used protocol analysis to validate a checklist method for
thought elicitation.  The second study described a prototype system for
automatically coding independent clauses into phrase categories (called
message frames), and the system's ability to classify new examples was
found to be limited.  The final study analyzed the performance of the
connectionist model.  The generalization performance of the model was
limited by the small number of examples, but an analysis of the receptive
fields of several message frames was consistent with previous research
about individual differences in reasoning about communication and
supported the assertion that individual linguistic forms are multi-
functional.  The model provided preliminary support for a situated,
phrasal lexical model of message design that does not rely on decontex-
tualized, functionally indexed abstractions.
 
Retrieve the file in the normal manner from archive.cis.ohio-state.edu.
Unfortunately, several of the figures are not included.  Hardcopy of the
figures will be sent to individuals who contact me directly by email.
 
And thanks of course to Jordan Pollack for providing what has become a
tremendously valuable resouce to the connectionist community.
 
Bruce Lambert
Department of Pharmacy Administration (M/C 871)
College of Pharmacy
University of Illinois at Chicago
833 S. Wood St.
Chicago, IL 60612
 
email: u53076 at uicvm.uic.edu