A Neuropsychological and Computational Investigation of Past Tense Verb Processing

将英语动词的现在时转换为过去时的简单过程主导了两种截然不同的语言知识观和加工观之间的争论。根据传统的符号理论，语言知识的形式是对离散的、符号表示进行操作的显性规则。相比之下，根据连接主义或神经网络理论，语言处理涉及大量类似神经元的处理单元的大规模并行相互作用。这项工作的主要目标是实现一个非连接主义风格的计算模型，该模型展示了语义和语音因素对过去时产生的不同影响，并反过来推导出可以通过未来神经心理测试进行评估的新预测。计算模拟将扩展Joanisse和Sedenberg的初步工作，在该工作中，动词加工涉及语音(理解和产生)的表征及其意义的平行交互作用。拟议的模拟将结合非现实主义的表示法和时间动力学。模型的初始阶段将采用固定长度的单音节单词的语音表征，但第二阶段将使用能够处理多音节项目的语音输入和输出的连续时间轨迹。语义表示将来自对大型文本语料库的一些分析，包括Lund和Burgess的超空间语言类比(HAL)，如有必要，将复制该模型，以便使生成的语义表示公开可供其他研究人员不受限制地使用。英语过去时结构中选择性损伤的理论理论性强的联结主义计算模型的建立，不仅对语义和语音表征在词汇加工中的作用提供了重要的见解，而且对语言系统中的主要成分及其相互作用的基本特征也提供了更广泛的认识。

The simple process of transforming the present tense of an English verb into its past-tense form has dominated the debate between two fundamentally different views of language knowledge and processing. According to traditional, symbolic theories, language knowledge takes the form of explicit rules operating over discrete, symbolic representations. By contrast, according to connectionist or neural-network theories, language processing involves the massively parallel interaction of large numbers of neuron-like processing units. The principal objective of the proposed work is to implement aconnectionist-style computational model that demonstrates the differentialinfluence of semantic and phonological factors on past-tense production and inturn, derives novel predictions that could be assessed by futureneuropsychological testing. The computational simulations will extend preliminary work byJoanisse and Seidenberg in which verb processing involved the parallelinteraction of representations of phonology (both in comprehension andproduction) and their meanings. The proposed simulations will incorporate morerealistic representations and temporal dynamics. An initial stage of modelingwill employ fixed-length phonological representations of monosyllabic verbstems, but a second stage will employ continuous-time trajectories forphonological input and output capable of handling multisyllabic items. Semanticrepresentations will be derived from a number of analyses of large textcorpora, including Lund and Burgess' Hyperspace Analogue to Language (HAL),which will be replicated if necessary in order to make the resulting semanticrepresentations publicly available for unrestricted use by other researchers. The development of a strongly theoretically motivated connectionist computational model of selective impairments in English past-tense formation will not only provide important insights into the role of semantic and phonological representations in lexical processing, but also more generally about the fundamental character of the principal components, and their interactions, in the language system.