Doctoral Dissertation Research: Sources of argument role insensitivity in verb processing

人类通常能快速而准确地理解话语，即使在嘈杂或退化的环境中听或读也是如此。许多研究人员将这一成功归因于人们快速预测即将到来的单词的能力。以前的研究已经证明了预测机制的各种证据，例如，预测的单词越多，阅读速度就越快。但人们对预测产生的机制知之甚少。这个项目通过关注人们似乎做出不恰当预测的情况来研究这些机制。一个有用的测试用例是“角色颠倒”句对，例如“服务员服务过的顾客”和“顾客服务过的服务员”，其中谁对谁做了什么就颠倒了。一些心理语言学的预测方法，特别是那些涉及理解的方法表明，动词“服务”在两个句子中都是一样的，尽管在第二个句子中是不合适的。这被认为是人类在产生期望时忽视了名词的作用的证据。然而，其他一些预测方法表明，人类充分利用角色信息，在相同的句子中产生适当的期望。该项目旨在解决这一差异。该项目结合计算和实验方法来调查为什么不同的测量表明语义角色在语言中的时刻预测中的作用更大或更小。该项目将开发一个语言预测的计算模型，试图捕捉一组共享的认知过程如何映射到不同的实验测量上。该模型将根据新实验的结果进行扩展。为了了解预测的时间进程和不同任务元素的贡献，实验将系统地改变参与者是否被显示出异常延续，以及参与者被要求做出什么样的反应。该项目还开发和完善了一个可扩展的管道，用于心理语言学实验中口语数据的半自动分析，可供其他研究人员使用。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。

Humans generally understand utterances quickly and accurately, even in noisy or degraded environments for listening or reading. Many researchers have attributed this success to people’s ability to rapidly predict upcoming words. Previous studies have demonstrated various kinds of evidence for prediction mechanisms, e.g., more predictable words are read more quickly. But less is known about the mechanisms by which predictions are generated. This project investigates these mechanisms, by focusing on situations where people appear to make inappropriate predictions. A useful test case is “role reversed” sentence pairs, such as “the customer that the waitress had served” and “the waitress that the customer served”, in which who did what to whom is reversed. Some psycholinguistic measures of prediction, particularly those involving comprehension, suggest that the verb “served” is equally expected in both sentences, despite being inappropriate in the second. This has been taken as evidence that humans ignore the roles of nouns when generating expectations. However, some other measures of prediction suggest that humans generate appropriate expectations in those same sentences, making full use of role information. This project seeks to resolve this discrepancy. The project combines computational and experimental methods to investigate why different measures indicate a greater or lesser role for semantic roles in moment-by-moment prediction in language. The project will develop a computational model of linguistic prediction that seeks to capture how a shared set of cognitive processes maps onto different experimental measures. The model will be extended based on results from new experiments. In order to understand the time course of predictions and the contributions of different task elements, the experiments will systematically vary whether or not participants are shown anomalous continuations, and what kind of response participants are required to give. The project also develops and refines a scalable pipeline for semi-automatic analysis of spoken language data in psycholinguistic experiments, which can be used by other researchers.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.