Collaborative Research: A Comparative Investigation of Statistical Learning

协作研究：统计学习的比较研究

• 批准号: 0519237
• 负责人: Daniel Weiss
• 金额: $ 22.66万
• 依托单位: Pennsylvania State Univ University Park
• 依托单位国家: 美国
• 项目类别: Continuing Grant
• 财政年份: 2005
• 资助国家: 美国
• 起止时间: 2005-09-15 至 2009-08-31
• 项目状态: 已结题
• 来源: https://www.nsf.gov/awardsearch/showAward?AWD_ID=0519237&HistoricalAwards=false
• 关键词: Collaborative; Research; Comparative; Investigation; Statistical

When adults listen to speech, their perception is influenced by the phonetic system of their native language. For example, Japanese speakers find it very difficult to hear the difference between English [r] and [l], while English speakers have trouble discriminating Hindi [d] and [D]. In contrast, infants who have not yet acquired the phonetic system of their native language often surpass adults in their ability to discriminate foreign contrasts. By 12 months, infants lose sensitivity to most foreign contrasts and show the adult-like pattern of discrimination. What is responsible for this change over the first year of life? With support from the National Science Foundation, Drs. Jessica Maye and Daniel Weiss will follow up on their research that has demonstrated that infants' sensitivity to statistical regularities in the speech they hear may lead to decreased sensitivity to foreign language contrasts as well as enhanced sensitivity to native language contrasts. They will measure the degree to which different distributional patterns in speech affect discrimination by human infants and adults as well as cotton-top tamarin monkeys. Previous research has found that tamarins, like humans, are capable of learning linguistic regularities through statistical properties of human speech. However, there are subtle differences across species in the nature of what is learned. This project will provide the first test of statistical phonetic learning in another species, with an aim to find both similarities and differences between humans and nonhumans, as well as between infants and adults. Furthermore, these experiments will involve the use of monkey call stimuli as well as human speech, in order to compare the species on their own vocalizations versus that of another species. Our ability to acquire language is one of the fundamental traits distinguishing humans from other animals. For hundreds of years, scholars have debated the extent to which the mechanisms underlying language acquisition are unique to humans, as it is clear that only humans ultimately acquire the full complexity of human language. One such learning mechanism is the ability to draw linguistic inferences on the basis of statistical patterns in the language that we hear. This is an ability that is not exclusive to humans; however, there appear to be differences in the units over which humans and nonhumans compute such regularities. Drs. Maye and Weiss hypothesize that these differences may underlie what makes humans such adept learners of human language. Similarly, there may be subtle differences in the statistical computations performed by infant and adult learners which would account for the fact that language acquisition that begins in infancy reaches a level of proficiency far surpassing that of languages learned later in life. This project will elucidate the similarities and differences between these three populations in their ability to learn statistical regularities in both speech stimuli and tamarin calls, in order to answer the question of what makes humans, and infants in particular, so remarkably adept at decoding an unknown language. On a practical level, the project will contribute to the infrastructure of science by training students in the interdisciplinary research of cognitive science (including linguistics, developmental psychology, and comparative psychology). Furthermore, a better understanding of what makes infants such successful language learners has far reaching clinical and technological applications. For example, a comparison of the learning mechanism of good learners (e.g. infants) vs. poor learners (e.g. adults, primates) may well lead to new therapies for the language disabled. Understanding how human listeners cope with enormous amounts of acoustic variation to uncover the message of a spoken utterance will help to improve speech recognition systems, which currently perform far below the level of human listeners.

当成年人听演讲时，他们的感知受到母语语音系统的影响。例如，说日语的人很难听出英语[r]和[l]之间的区别，而说英语的人很难区分印地语[d]和[D]。相比之下，还没有掌握母语语音系统的婴儿在辨别外来对比的能力上往往超过成年人。到12个月大时，婴儿对大多数外来对比失去了敏感性，并表现出类似成年人的辨别模式。是什么导致了生命第一年的这种变化？在国家科学基金会的支持下，Jessica Maye博士和丹尼尔韦斯博士将继续他们的研究，该研究表明，婴儿对他们听到的语音中的统计差异的敏感性可能会导致对外语对比的敏感性降低，以及对母语对比的敏感性增强。他们将测量不同的语言分布模式对人类婴儿和成年人以及棉顶绢毛猴的辨别力的影响程度。先前的研究发现，绢毛猴和人类一样，能够通过人类语言的统计特性来学习语言。然而，在学习的本质上，不同物种之间存在着微妙的差异。该项目将提供另一个物种的统计语音学习的第一个测试，旨在发现人类和非人类之间以及婴儿和成人之间的相似性和差异。此外，这些实验将涉及使用猴子呼叫刺激以及人类语音，以比较物种对自己的发声与另一个物种。我们获得语言的能力是人类区别于其他动物的基本特征之一。数百年来，学者们一直在争论语言习得的机制在多大程度上是人类独有的，因为很明显，只有人类最终获得了人类语言的全部复杂性。其中一种学习机制是根据我们所听到的语言中的统计模式进行语言推理的能力。 这种能力并不是人类独有的;然而，人类和非人类计算这种能力的单位似乎存在差异。Maye博士和韦斯博士假设，这些差异可能是人类如此熟练地学习人类语言的原因。同样，婴儿和成人学习者进行的统计计算可能存在细微的差异，这可以解释婴儿期开始的语言习得达到的熟练程度远远超过以后学习的语言。这个项目将阐明这三个种群在语音刺激和绢毛猴叫声中学习统计学能力的异同，以回答是什么使人类，特别是婴儿，如此擅长解码未知语言的问题。在实践层面上，该项目将通过培训学生进行认知科学（包括语言学，发展心理学和比较心理学）的跨学科研究来促进科学的基础设施。此外，更好地了解是什么使婴儿如此成功的语言学习者具有深远的临床和技术应用。例如，比较好的学习者（如婴儿）与差的学习者（如成人、灵长类动物）的学习机制，很可能会为语言障碍者带来新的治疗方法。理解人类听众如何科普大量的声学变化以揭示口语话语的信息将有助于改进语音识别系统，目前语音识别系统的性能远远低于人类听众的水平。