Networks and Morphophonemic Rules Revisited

In the debate over the power of connectionist models to handle linguistic phenomena, considerable attention has been focused on the learning of simple morphophonemic rules. Rumelhart and McClelland’s celebrated model of the acquisition of the English past tense (1986), which used a simple pattern associator to learn mappings from stems to past tense forms, was advanced as evidence that networks could learn to emulate rule-like linguistic behavior. Pinker and Prince’s equally celebrated critique of the past-tense model (1988) argued forcefully that the model was inadequate on several grounds. For our purposes, these are (1)ּthe fact that the model is not constrained in ways that humans language learners clearly are and (2)ּthe fact that, since the model cannot represent the notion “word”, it cannot distinguish homophonous verbs. A further deficiency of the model, one not brought out by Pinker and Prince, is that it is not a processing account: the task that the network learns is that of associating forms with forms rather than that of producing forms given meanings or meanings given forms. This paper describes a connectionist model which addresses all three objections to the earlier work on morphophonemic rule acquisition. The model learns to generate forms in one or another “tense”, given arbitrary patterns representing “meanings”, and to output the appropriate tense given forms. The inclusion of meanings in the network means that homophonous forms are distinguished. In addition, this network experiences difficulty learning reversal processes which do not occur in human language.