| The relationship between learning and processing is a complex problem for theories of language. A connectionist model of sentence production is used to instantiate the relationship between these two domains. The model addresses the issue of how learning can be both sensitive to regularities in the data, and also able to generalize to novel situations (Chapter 2). To accomplish this, the model uses an architecture with separate pathways for lexical selection and sequencing (Dual-path architecture). This Dual-path model makes use of role-concept variables that are inspired by distinctions in the visual system (Landau & Jackendoff, 1993). This Dual-path architecture generalizes symbolically better than other architectures, provides an account of the relationship between symbolic and statistical mechanisms in verb generalization during acquisition, and models double dissociations that occur in aphasia. In chapter 3, a structural priming study on humans is performed to test the assumption of this model (and many linguistic theories) that event roles are important in sentence production. In chapter 4, the Dual-path model is used to model structural priming results. The model explains structural priming as the same type of implicit learning that the model used to learn the grammar in the first place. The model's ability to account for structural priming depended both on the architecture and on the role representation that it used to represent the message. In chapter 5, a simpler version of the model is used to examine how the back-propagation learning algorithm constrains the learning of syntactic representations. Several mini-grammars were compared, using a new technique for mapping weights in a network (labelprop ). This comparison demonstrated how representations in the model come out of the interaction of the architecture and the properties of the grammar. |
