TR on language acquisition

[Michael Gasser]

FTP-host: cs.indiana.edu (129.79.254.191)
FTP-filename: /pub/techreports/TR384.ps.Z

The following paper is available in compressed postscript form by
anonymous ftp from the Indiana University Computer Science Department
ftp archive (see above).  The paper is 60 pages long.  Hardcopies
won't be available till September, I'm afraid.

Comments welcome.

Michael Gasser
gasser at cs.indiana.edu

=================================================================
		       Learning Words in Time:
	       Towards a Modular Connectionist Account
	      of the Acquisition of Receptive Morphology

			    Michael Gasser
	     Computer Science and Linguistics Departments
			  Indiana University

   To have learned the morphology of a natural language is to have the
capacity  both to  recognize and to produce  words consisting of novel
combinations   of  familiar  morphemes.   Most   recent  work  on  the
acquisition of morphology takes the perspective of  production, but it
is receptive  morphology which  comes  first in the child.  This paper
presents  a connectionist model of the  acquisition of the capacity to
recognize morphologically complex words.  The model takes sequences of
phonetic  segments  as  inputs  and  maps   them   onto  output  units
representing  the meanings of  lexical  and grammatical morphemes.  It
consists  of  a simple recurrent  network  with  separate hidden-layer
modules for  the tasks of recognizing  the root  and  the  grammatical
morphemes of the  input  word.  Experiments  with artificial  language
stimuli demonstrate  that the model generalizes  to  novel  words  for
morphological rules of all but one of the major types found in natural
languages  and  that  a  version   of  the  network   with  unassigned
hidden-layer   modules  can  learn  to  assign  them   to  the  output
recognition tasks in an efficient manner.  I also argue that for rules
involving reduplication, that is, the copying of portions  of  a root,
the  network requires separate recurrent subnetworks for  sequences of
larger units such as syllables.  The network can learn to develop  its
own syllable representations which not only support the recognition of
reduplication but also provide the basis for  learning to produce,  as
well as  recognize,  morphologically  complex words.  The  model makes
many detailed predictions about  the learning difficulty of particular
morphological rules.