A new approach to computational model evaluation: Benchmarking against infant experimental data

这个项目是关于确定学习英语和西班牙语的婴儿何时对母语中辅音和元音的顺序变得敏感。例如，英语单词可以以“st”开头(如“State”或“Stamp”)，而西班牙语单词则不能。研究人员将使用婴儿研究的实验结果来评估计算模型做出的预测，这些模型代表了关于婴儿如何学习语言的不同假设。这一比较将提供对婴儿使用的具体机制的洞察。此外，揭示婴儿何时收听声音排序限制的详细时间表将使人们能够更好地了解一些患有沟通障碍的儿童在学习声音序列方面遇到困难的根源。该项目的第二个成果是将讲西班牙语的人口更多地纳入发展研究，这一点在美国不断变化的人口结构下非常重要。近年来，许多计算模型被提出来解释婴儿是如何学习语言的。当以成人说话者的表现为基准时，多个模型架构通常会获得相似的成功水平。在这个项目中，计算模型将直接针对婴儿的行为进行评估，以便为区分它们提供额外的标准。更具体地说，研究人员将评估婴儿对母语中声音顺序限制的知识是否可以得到不同计算模型的分词输出的支持。除了提供获得合理测序限制的详细时间表和更好地了解婴儿使用的发育机制外，该项目还将作为未来研究的模板，旨在将计算模型直接与婴儿行为进行比较。该奖项反映了NSF的法定使命，并通过使用基金会的智力优势和更广泛的影响审查标准进行评估，被认为值得支持。

This project is about determining when English- and Spanish-learning infants become sensitive to the sequencing of consonants and vowels in their native language. For example, English words can begin with ‘st’ (as in ‘state’ or ‘stamp’), while Spanish words cannot. The researchers will use experimental findings from studies on infants to evaluate predictions made by computational models that represent different hypotheses about how infants learn language. This comparison will provide insight into the specific mechanisms used by infants. In addition, uncovering the detailed timeline of when infants tune into sound sequencing restrictions will enable a better understand the roots of the difficulties some children with communication disorders have in learning sound sequencing. A secondary outcome of this project is the greater inclusion of Spanish-speaking populations in developmental research, which is important given the changing demographics in the United States.In recent years, many computational models have been proposed to explain how infants learn language. Multiple model architectures typically achieve similar levels of success when benchmarked against the performance of adult speakers. In this project, computational models will be evaluated directly against infant behavior in order to provide additional criteria for distinguishing between them. More specifically, the researchers will evaluate whether infants’ knowledge of sound sequencing restrictions in their native language can be supported by the output of different computational models of word segmentation. In addition to providing a detailed timeline of the acquisition of sound sequencing restrictions and a better understanding of the developmental mechanisms used by infants, this project will also serve as a template for future studies that aim to compare computational models directly against infant behavior.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.