Integration of Connectionist and Symbolic Computation for Linguistic Modeling

这是一个为期三年的持续资助的第一年， 授予IRI-9213894。 在以前的研究中， 语言研究，和声语法，依赖于一种新颖的方法， 整合了联结主义和符号计算 当前 研究这种方法的扩展，关于正式 自然语言的语法和形式语言的语法。 具体地说，以前的工作是扩展：包括 递归的、完全分布式的联结主义表示; 开发一种显式符号编程语言来描述 一个新的计算器对这些表示进行计算， 类连接网络;研究的表达能力 调和语法描述形式语言，从一个 上下文无关语言的研究;分析 连接主义网络导致了一个递归的，嵌入式的- 不变量良构（Harmony）函数;测试能力 调和语法理论对语言学问题的处理至关重要 包括嵌入;评估我们充分 用于编码特征的分布式表示 语言表征中的句法/语义结构角色; 扩展理论，将符号计算 统一的概念，并将其应用于基于统一的 自然语言语法的方法;研究有效性 和时间行为的连接主义解析器，实现了这两个 调和文法的数值和非数值形式化; 并扩展了最近关于符号联结学习的研究， 规则系统到谐波语法的特殊情况。//

This is the first-year funding of a three-year continuing award IRI-9213894. In previous research, an approach to the formal study of language, Harmonic Grammar, relies on a novel means of integrating connectionist and symbolic computation. The current research studies extensions of this approach, concerning formal grammars of natural languages and grammars of formal languages. Specifically, the previous work is extended: to encompass recursive, fully distributed connectionist representations; to develop an explicit symbolic programming language to characterize the computations over these representations carried out by a new class of connectionist networks; to study the expressive power of Harmonic Grammar to describe formal languages, starting with a study of context-free languages; to analyze the properties of connectionist networks that lead to a recursive, embedding- invariant well-formedness (Harmony) function; to test the ability of Harmonic Grammar to deal with linguistic problems crucially involving embedding; to assess the effectiveness of our fully distributed representations for encoding the features of syntactic/semantic structural roles in linguistic representations; to extend the theory to incorporate the symbolic computational notion of unification, and to apply it to unification-based approaches to natural language grammar; to study the effectiveness and temporal behavior of connectionist parsers that implement both the numerical and nonnumerical formalizations of Harmonic Grammar; and to extend recent work on connectionist learning of symbolic rule systems to the particular case of Harmonic Grammar.//