Learning Morphophonemic Processes without Underlying Representations and Explicit Rules

Traditional phonology presupposes abstract underlying representations (UR) and a set of rules to explain the phonological phenomena. There are, however, a number of questions that have been raised regarding this approach : Where do URs come from? How are rules found and related to each other? In the current study, a connectionist network was trained without the benefit of any UR and explicit rules. We hypothesized that rules would emerge as the generalizations the network abstracts in the process of learning to associate forms -(sequences of phonological segments comprising words) with meanings (of the words) and URs as a pattern on the hidden layer. Employing a simple recurrent network we ran a series of simulations on different types of morphophonemic processes. The results of the simulations show that this network is capable of learning morphophonemic processes without any URs and explicit rules.