Verb Learning and The Early Development of Sentence Comprehension: Experimental and Computational Studies

在NSF的支持下，Cynthia Fisher博士、Dan Roth博士和Yael Gertner博士将对句子理解的早期发展进行为期三年的研究。句子理解中的一项基本任务是将语义角色分配给句子成分，从而决定谁对谁做了什么。句法自举理论认为，即使是非常年幼的儿童也会使用成人句法知识的前身来完成这一任务。该项目结合了实验和计算方法来测试和完善这一理论。这项实验工作将调查15至30个月大的婴儿对所提出的简单句法线索对句子解释的敏感性。计算工作涉及一个自动语义角色标注系统；语义角色标注的计算模型学习识别填充语义角色的句子成分，并确定它们的角色，如施动者或目标。这个项目的两个组成部分将检验句子结构的简单方面指导早期句子理解的提议，通过检查儿童对句子理解的成功和告诉人的错误，通过基于所提出的句子结构的浅层描述的明确的学习建模，通过比较计算模型的学习进度和错误与儿童的学习进度和错误。这种实验和计算研究的结合既是为了促进关于儿童如何学习他们的母语的科学知识，也是为了指导新的、强大的学习协议的发展，这些协议将用于自动自然语言处理。

With NSF support, Dr. Cynthia Fisher, Dr. Dan Roth, and Dr. Yael Gertner will conduct a three-year study of the early development of sentence comprehension. A fundamental task in sentence comprehension involves assigning semantic roles to sentence constituents, thus determining "who does what to whom". The syntactic bootstrapping theory proposes that even very young children use precursors of the adult's knowledge of syntax to accomplish this task. The project combines experimental and computational approaches to test and refine this theory. The experimental work will investigate 15- to 30-month-olds' sensitivity to proposed simple syntactic cues to sentence interpretation. The computational work involves a system for automatic semantic role labeling; computational models of semantic role labeling learn to identify sentence constituents that fill a semantic role, and to determine their roles, such as agent or goal. The two components of this project will test the proposal that simple aspects of sentence structure guide early sentence comprehension, by examining both successes and tell-tale errors in children's comprehension of sentences, by explicitly modeling learning based on the proposed shallow description of sentence structure, and by comparing the computational model's learning progress and errors to those of the children. This combination of experimental and computational studies is intended both to advance scientific knowledge about how children learn their native languages, and to guide the development of new, robust learning protocols that will be of use in automatic natural language processing.