chapter  4
Language Acquisition in a Self-Organising Neural Network Model
WithPing Li
Pages 31

One crucial aspect of human language learning is the learner’s ability to generalise existing patterns to novel instances. This ability often leads to various erroneous generalisations in learning. “Overgeneralisation” is one such type of error, characterised by the learner’s use of a linguistic pattern that is broader in scope than the corresponding adult uses (Bowerman, 1982; Brown, 1973; Clark, 1987; Pinker, 1989). Perhaps the bestknown example of overgeneralisation is the acquisition of the English past tense: children generalise -ed to irregular verbs, producing errors like falled, breaked, and comed (Brown, 1973; Kuczaj, 1977). But just what leads to children’s overgeneralisations has been under intensive debate. Using the acquisition of the English past tense as an example, researchers have debated whether language acquisition should be characterised as a symbolic, rule-based learning process or as a connectionist, statistical learning process. Symbolic theorists assume that overgeneralisation errors result from the child’s internalisation and application of linguistic rules (Ling & Marinov, 1993; Marcus, Pinker, Ullman, Hollander, Rosen & Xu, 1992; Pinker, 1991, 1999; Pinker & Prince, 1988), whereas connectionists argue that overgeneralisations reflect the child’s ability to extract statistical regularities from the input (MacWhinney & Leinbach, 1991; Plunkett & Marchman, 1991, 1993; Rumelhart & McClelland, 1986; Seidenberg, 1997).