tech report available

[Michael Gasser]

                    NETWORKS THAT LEARN PHONOLOGY

                           Michael Gasser
                            Chan-Do Lee
                     Computer Science Department
                         Indiana University
                        Bloomington, IN 47405

                         Technical Report 300
                            December 1989

Abstract:

Natural language phonology presents a challenge to connectionists
because it is an example of apparently symbolic, rule-governed
behavior.  This paper describes two experiments investigating the
power of simple recurrent networks (SRNs) to acquire aspects of
phonological regularity.  The first experiment demonstrates the
ability of an SRN to learn harmony constraints, restrictions on the
cooccurrence of particular types of segments within a word.  The
second experiment shows that an SRN is capable of learning the kinds
of phonological alternations that appear at morpheme boundaries, in
this case those occurring in the regular plural forms of English nouns.
This behavior is usually characterized in terms of a derivation from a
more to a less abstract level, and in previous connectionist treatments
(Rumelhart & McClelland, 1986; Plunkett & Marchman, 1989) it has been
dealt with as a process of yielding the combined form (plural) from the 
simpler form (stem).  Here the behavior takes the form of the more 
psychologically plausible process of the production of a sequence of 
segments given a meaning or of a meaning given a sequence of segments.
This is accomplished by having both segmental and semantic inputs and
outputs in the network.  The network is trained to auto-associate the
current segment and the meaning and to predict the next phoneme.

				   
----------------------------------------------------------------------

To order copies of this tech report, send  mail to Nancy Garrett at
nlg at cs.indiana.edu / Computer Science Department, Indiana University,
Bloomington, IN 47405.