CAREER: Modeling Linguistic Competence and Performance in a Hybrid Connectionist Framework

自然语言理解（NLU）系统的一个核心方面是创建和操纵编码语法知识的复杂符号结构。 然而，同样重要的是， 衡量数字影响的能力，如词汇频率，它决定了对歧义的首选解释 话语。 的传统方法 NLU 强调语言的符号方面，而联结主义模型则侧重于数字方面。 在这项研究中，一个混合连接主义框架的开发，充分整合了两种类型的知识在一个统一的计算方法。 在一个实现的原型，符号特征传递编码丰富的语法结构，最近的语言学理论，而竞争性的数字处理捕捉的影响，消除歧义的话语。 这个项目涉及到一些扩展的模型，需要显着的发展，在一个并行的网络架构内的符号和数字处理的集成。 在分布式符号处理和复杂竞争动态的控制中探索新的算法，特别关注单个词汇的结构和数值影响。 信息. 研究还 调查 假设可以从大型文本语料库中提取准确的数字偏好信息。

A central aspect of a natural language understanding (NLU) system is the creation and manipulation of complex symbolic structures that encode grammatical knowledge. However, equally important is the ability to weigh numeric influences, such as lexical frequencies, that determine the preferred interpretation of an ambiguous utterance. Traditional approaches to NLU have emphasized the symbolic aspects of language, while connectionist models have focused on the numeric aspects. In this research, a hybrid connectionist framework is developed that fully integrates the two types of knowledge within a uniform computational approach. Within an implemented prototype, symbolic feature passing encodes the rich grammatical structures of a recent linguistic theory, while competitive numeric processing captures the influences that disambiguate an utterance. This project involves a number of extensions to the model that necessitate significant developments in the integration of symbolic and numeric processing within a parallel network architecture. Novel algorithms are explored within distributed symbolic processing and the control of complex competitive dynamics, especially focusing on the structural and numeric influences of individual lexical information. The research also investigates the hypothesis that accurate numeric preference information can be extracted from large text corpora.