A connectionist model of the Arabic plural system

Connectionist accounts of inflectional morphology have focused on the domain of the English past tense in which the default process (add/ed/) reflects the process of suffixation adopted by the majority of forms in the language. Connectionist models exploit the imbalance between English regular and irregular verbs when learning the past tense and when responding to novel forms in a default fashion. Not all inflectional systems have a default which is characterised by a majority of forms in the language. The Arabic plural system has been cited as one such system where a minority default process operates-the "sound" plural in Arabic applies to only a minority of the forms in the lexicon, yet it appears to adopt the role of a default for novel nouns. We contrast two types of default process that might lead to this behaviour, a symbolic default and a distributional default. We provide a detailed analysis of the phonological similarity structure of the Arabic plural system and conclude that it is unlikely the distribution of singular Arabic nouns can support a distributional default for the sound plural. Nevertheless, classification networks and pattern association networks perform surprisingly well at the task of learning the Arabic plural and generalising to unseen Arabic nouns. In fact, the performance of a neural network classifier exceeds the performance of a symbolic hybrid "rule-associative" model even when the latter is optimised for correct classification. We conclude that the default rule is redundant, since generalisation to novel forms is more accurately predicted by the similarity structure in the phonological space of Arabic nouns alone. Finally, we present a set of empirical predictions for children's; acquisition of the Arabic plural system, derived from the network simulations.