The nature of the semantic representations of words, and how these representations interact with other forms of lexical knowledge, is fundamental to linguistic and cognitive behavior. The proposed research employs computational modeling to provide insight into lexical semantic processing, both in normal individuals and in neurological patients with semantic impairments due to brain damage, and into the nature of recovery after semantic damage. The research is cast within a connectionist/parallel distributed processing (PDP) framework in which groups of simple, neuron-like processing units representing different types of lexical information (e.g., orthographic, phonological, semantic) interact across weighted connections in the process of settling into stable patterns of activity representing the network's interpretation of a given input. This framework provides a rich set of computational principles that provide new ways of thinking about how cognitive processes are implemented in the brain. The computational principles central to the proposed research are that learning and processing, both in the normal system and after damage, are sensitive to the similarity of representations within a domain, to the frequency and consistency of the mappings between domains, and to the temporal contiguity of representations both within and between domains. Three sets of simulations are proposed. The first set (1-4) addresses normal lexical semantic processing as reflected by semantic and associative priming in naming and lexical decision. The models encode semantic relatedness by pattern overlap and associative relatedness by temporal contiguity during training, and account for empirical phenomena relating to interactions between types of priming, priming across unrelated items, and lexical ambiguity resolution. The second set (5-8) investigates the degree of specialization within semantics, as evidenced by the occurrence of selective deficits following brain damage, by exploring the effects of damage to a system in which internal representations--corresponding to lexical semantics--must learn to mediate between multiple modalities in performing lexical tasks. The issues addressed include access vs. degraded-store deficits, modality-specific aphasias, category-specific impairments, and abnormal priming in dementing patients. The third set (9-12) investigates generalized recovery via retraining after damage and explores the benefits of retraining strategies employing multiple modalities and tasks. The research provides new insights into normal and impaired lexical semantic processing and should contribute to improved patient therapy.
