Statistical approaches to linguistic pattern learning

语言模式学习的统计方法

• 批准号: 8511737
• 负责人: Richard N. Aslin
• 金额: $ 28.91万
• 依托单位: UNIVERSITY OF ROCHESTER
• 依托单位国家: 美国
• 财政年份: 1999
• 资助国家: 美国
• 起止时间: 1999-02-01 至 2015-05-31
• 项目状态: 已结题
• 来源: https://reporter.nih.gov/project-details/8511737
• 关键词: Accounting; Address; Adult; Affect; Age; Auditory; Benchmarking; Categories; Child; Complex; Cues; Developmental Delay Disorders; Disease; Elements; Event; Exposure to; Eye Movements; Feedback; Frequencies; Goals; Health; Hearing; Hour; Human; Human Development; Indium; Infant; Instruction; Judgment; Language; Language Delays; Language Development; Language Disorders; Lead; Learning; Legal; Linguistics; Machine Learning; Maps; Methods; Nature; Noise; Participant; Pattern; Performance; Phase; Play; Positioning Attribute; Process; Production; Property; Reaction Time; Recurrence; Relative (related person); Research; Resources; Role; Sampling; Series; Shapes; Specificity; Structure; Techniques; Testing; Time; Training; Ursidae Family; Variant; Visual; design; lexical; natural language; novel; programs; public health relevance; remediation; research study; response; scale up; sound; statistics; visual motor

DESCRIPTION (provided by applicant): The purpose of the proposed research is to provide a comprehensive account of the factors that affect how infants, children, and adults learn the categories of their native language from distributional information in linguistic input. The categories of a language consist of sets of words (e.g., noun, verb) that play a functionally equivalent role in grammatical sentences. Distributional information refers to the patterning of elements in a large corpus of sentences and includes how frequently those elements occur, what position they occupy in a sentence, and the context provided by neighboring elements. Our longstanding program of research on statistical learning in word segmentation (how learners determine which sound sequences form words) has documented the power, rapidity, and robustness of infants, children, and adults sensitivity to complex distributional information. Here we extend that program of research to a crucial aspect of learning higher-level structures of language. In our proposed studies, we use a miniature artificial language paradigm that affords us complete control over all the distributional cues in the input, something that is virtually impossible using real languages. Participants listen to a sample of utterances and make judgments about their acceptability. Crucially, during a learning phase, they do not hear all possible utterances that are "legal" in the artificial language; some are withheld for use in a later post-test. The post-test utterances either conform to the distributional patterns present in the learning phase, or they violate those patterns. The key test is whether participants judge novel-but-legal utterances to be acceptable, thereby showing the ability to generalize correctly beyond the input to which they were exposed. Studies of children provide additional support for learning the distributional cues by pairing utterances with videos of simple events. Studies of adults will be used for comparison, and will also present them with learning materials in the visual-motor domain to assess the detailed time-course of learning and the specificity of the results to auditory linguistic materials. Taken together, the results of these studies of infants, children, and adults will document the key structural variables in language learning that enable a distributional mechanism of category formation to operate and will highlight the ways these mechanisms may differ over age and domain. PUBLIC HEALTH RELEVANCE: Language development is one of the hallmarks of the human species, yet it is difficult to study because of the huge variation in early exposure to different amounts of linguistic input. The use of artificial languages that are acquired in the lab over a few hours provides a window on the mechanisms of language development. We will study language learning in the lab to gain a unique perspective on how the categories (noun, verb, etc) are formed from listening to the patterns of words in a small set of sentences. These studies will not only reveal a basic mechanism of language learning, but also establish benchmarks against which language delay can be compared. Moreover, understanding the mechanisms that lead to successful acquisition in normal children can help to identify loci of language disorders and design methods for remediating disorders.

描述（由申请人提供）：拟议研究的目的是全面说明影响婴儿、儿童和成人如何从语言输入的分布信息中学习母语类别的因素。语言的类别由在语法句子中发挥功能等效作用的单词集（例如名词、动词）组成。分布信息是指大型句子语料库中元素的模式，包括这些元素出现的频率、它们在句子中占据的位置以及相邻元素提供的上下文。我们对分词统计学习（学习者如何确定哪些声音序列构成单词）的长期研究计划记录了婴儿、儿童和成人对复杂分布信息的敏感性的力量、快速性和鲁棒性。在这里，我们将该研究计划扩展到学习高级语言结构的一个关键方面。在我们提出的研究中，我们使用微型人工语言范式，使我们能够完全控制输入中的所有分布线索，这在使用真实语言时几乎是不可能的。参与者听取话语样本并对其可接受性做出判断。至关重要的是，在学习阶段，他们不会听到人工语言中所有可能的“合法”话语；有些被保留以供稍后的后测试使用。测试后的话语要么符合学习阶段存在的分布模式，要么违反这些模式。关键的测试是参与者是否判断新颖但合法的话语是可以接受的，从而显示出正确概括超出他们所接触到的输入的能力。对儿童的研究通过将言语与简单事件的视频配对来为学习分布线索提供额外的支持。对成人的研究将用于比较，并且还将向他们提供视觉运动领域的学习材料，以评估详细的学习时间过程以及结果对听觉语言材料的特异性。总而言之，这些针对婴儿、儿童和成人的研究结果将记录语言学习中的关键结构变量，这些变量使类别形成的分配机制能够发挥作用，并将强调这些机制随着年龄和领域的不同而可能存在的差异。公共健康相关性：语言发展是人类的标志之一，但由于早期接触不同数量的语言输入存在巨大差异，因此研究起来很困难。使用在实验室几个小时内获得的人工语言为了解语言发展机制提供了一个窗口。我们将在实验室中研究语言学习，以获得关于如何通过聆听一小组句子中的单词模式来形成类别（名词、动词等）的独特视角。这些研究不仅将揭示语言学习的基本机制，还将建立比较语言延迟的基准。此外，了解正常儿童成功习得的机制有助于识别语言障碍的位点并设计治疗障碍的方法。