Semantic Representations in Reading

连接主义网络通过大规模并行的合作和竞争交互作用计算许多简单的神经元样处理单元，为研究认知过程及其神经实现提供了一个统一的框架。这项拟议的研究涉及通过开发和“损害”正常阅读的连接主义模型来对脑损伤患者的后天阅读障碍进行建模。基于心理学和神经心理学的数据，该模型有两种方法来推导印刷单词的发音：一种是通过“试探”单词从正字法直接映射到音系学的语音路径；另一种是通过“识别”单词的意思通过语义间接映射的语义路径。隔离地对这些路径中的每一个进行初步的连接主义实现已经成功地再现了有限范围的正常和异常阅读行为。建议开发一个完整的双路径模型，能够处理更广泛的现象，涉及阅读习得、成熟技能阅读、部分损伤后的行为和恢复过程中每条路径的协调和相对贡献。该模型将在神经生理学上更可信的连接主义形式主义中开发，而不是在之前的工作中使用的。此外，一个主要的研究工作将集中在开发一个足够复杂的语义系统，以克服现有系统的严重限制。将使用高性能计算设施来满足训练和测试该模型的广泛计算需求。

Connectionist networks, which compute via the massively parallel cooperative and competitive interactions many simple neuron-like processing units, provide a unified framework for studying cognitive processes and their neurological implementation. The proposed research involve modeling acquired reading disorders in patients with brain damage by developing and "lesioning" a connectionist model of normal reading. Based on data from psychology and neuropsychology, the model has two routes for deriving the pronunciation of a printed word: a phonological route that maps directly from orthography to phonology by "sounding-out" the word, and a semantic route that maps indirectly via semantics by "recognizing" the meaning of the word. Preliminary connectionist implementations of each of these routes in isolation have succeeded in reproducing a limited range of normal and abnormal reading behavior. It is proposed to develop a complete dual-route model capable of addressing a much wider range of phenomena, relating to the coordination and relative contribution of each route during reading acquisition, mature skilled reading, behavior after partial damage, and recovery from damage. The model will be developed within a more neurophysiologically plausible connectionist formalism than the one used in the previous work. In addition, a major research effort will focus on developing an adequately sophisticated semantic system that overcomes the severe limitations of existing systems. High-performance computing facilities will be used to meet the extensive computational demands of training and testing the model.