Phonological constraint induction in a connectionist network: learning OCP-Place constraints from data

A significant problem in computational language learning is that of inferring the content of well-formedness constraints from input data. In this article, we approach the constraint induction problem as the gradual adjustment of subsymbolic constraints in a connectionist network. In particular, we develop a multi-layer feed-forward network that learns the constraints that underlie restrictions against homorganic consonants, or ‘OCP-Place constraints’, in Arabic roots. The network is trained using standard learning procedures in connection science with a representative sample of Arabic roots. The trained network is shown to classify actual and novel Arabic roots in ways that are qualitatively parallel to a psycholinguistic study of Arabic. Statistical analysis of network behavior also shows that activations of nodes in the hidden layer correspond well with violations of symbolic well-formedness constraints familiar from generative phonology. In sum, it is shown that at least some constraints operative in phonotactic grammar can be learned from data and do not have to be stipulated in advance of learning.