Gaining a greater understanding of the mechanisms by which we acquire and process language is a critical first step in the development of strategies to aid those with acquired or developmental language deficits. Many direct methods for investigating the neural basis of language are unavailable because, unlike other domains of cognitive psychology such as vision or memory, language has no real non-human correlate. Therefore, we must rely more heavily on computational models, in conjunction with behavioral measures, to induce the workings of the language system. Connectionist models, in particular, offer the potential to provide novel explanations for many critical questions in the study of language, including the means by which children are able to learn language in noisy environments with limited feedback, how learning of a second language differs in adulthood and childhood, and how semantic, pragmatic, and discourse information is rapidly integrated with syntactic knowledge during comprehension. But connectionist models have so far been applied to only a limited number of the many complex aspects of language. The main goal of this proposal is to further develop, refine, and evaluate a connectionist model aimed at accounting for a broad range of sentence processing phenomena. Predictions of the model will be tested using self-paced reading-time and possibly eye-tracking measures. The model will also be studied to determine the general principles that characterize its behavior. Additional parts of the proposed project will study the induction of word meaning from large text corpora and explore the effects of experience on second language learning in connectionist networks.
