ABSTRACT
This chapter discusses interfaces in connectionist phonology. Connectionist theory was developed initially by cognitive psychologists, with a focus on problems and issues other than phonology per se. By the late 1980s, the major focus of connectionist modeling shifted to parallel distributed processing models, which learned the target phenomena from scratch and which contained some nodes that did not map easily onto symbolic elements. Rumelhart and McClelland developed the first nonrecurrent distributed model to address morphology. The input was a distributed representation of base word's pronunciation. Regular morphological patterns are generally associated with hundreds of lexical items, and so have enough base support for ready generalization, leading even to over-regularization of irregular past-tense forms. Distributed models of morphology have attracted greatest attention, because of their learning component. The chapter addresses phonological effects in language production that derive from error-driven learning. Errors are more likely to involve whole segments than individual features. This derives from the nature of feedback in local models.
