Building a comprehensive theory of pragmatic language through large-scale experiments, computation, and neurodiversity

该奖项是美国国家科学基金会社会、行为和经济科学博士后研究奖学金（SPRF）计划的一部分。SPRF计划的目标是为学术界、工业界或私营部门和政府的科学事业准备有前途的早期职业博士级科学家。SPRF奖励包括在知名科学家的赞助下进行为期两年的培训，并鼓励博士后进行独立研究。美国国家科学基金会寻求促进科学界各阶层的科学家，包括那些未被充分代表的群体的科学家，参与其研究项目和活动；博士后阶段被认为是实现这一目标的一个重要的专业发展阶段。每个博士后必须解决各自学科领域的重要科学问题。在Edward Gibson博士和Evelina Fedorenko博士的赞助下，这个博士后奖学金奖支持一位研究人类和机器如何理解非字面语言的早期职业科学家。人类的大部分交流都不是直接用语言编码的：我们可能会说“时间不早了”，礼貌地表示我们想要离开，或者称芭蕾舞演员为“天鹅”，以捕捉她的优雅。这种语言被称为语用学，通常被认为包括幽默、善意的谎言、隐喻、含意、韵律等等。虽然已经对这些现象进行了单独的调查，但尚不清楚它们是否受到相同机制的支持，也不知道它们在个体内部是如何联系的，这对面临沟通挑战的神经分化个体有影响。而且，虽然目前的语言模型显示出令人印象深刻的结果，但很少有研究探索哪种计算可能支持实用的语言理解，这对人工智能的成功至关重要。通过行为和计算方法，当前的项目将通过识别人类相关的语用推断集群并探索它们如何在模型中计算来解决这些限制。这个项目采用新的方法来阐明语用语言的统一框架，目标是在这个过程中包括神经分化个体作为研究人员。大规模的个体差异研究将发展出一整套实用语言，并揭示能力之间的关系（例如，理解善意的谎言与理解讽刺的关系最好吗？）非语言认知评估将确定哪些集群与社会推理、字面语言理解和执行能力有关。通过评估非字面语言解释的当前语言模型，该项目还揭示了语言模型在语用语言上的表现的可学习性、范围和泛化，并将使我们能够将模型活动与人类研究中发现的群集进行比较。通过对人类语用语言和现有计算模型的测试，本研究将使我们更接近于理解语用语言理解所需的输入和计算。该奖项反映了美国国家科学基金会的法定使命，并通过使用基金会的知识价值和更广泛的影响审查标准进行评估，被认为值得支持。

This award was provided as part of NSF's Social, Behavioral and Economic Sciences Postdoctoral Research Fellowships (SPRF) program. The goal of the SPRF program is to prepare promising, early career doctoral-level scientists for scientific careers in academia, industry or private sector, and government. SPRF awards involve two years of training under the sponsorship of established scientists and encourage Postdoctoral Fellows to perform independent research. NSF seeks to promote the participation of scientists from all segments of the scientific community, including those from underrepresented groups, in its research programs and activities; the postdoctoral period is considered to be an important level of professional development in attaining this goal. Each Postdoctoral Fellow must address important scientific questions that advance their respective disciplinary fields. Under the sponsorship of Dr. Edward Gibson and Dr. Evelina Fedorenko, this postdoctoral fellowship award supports an early career scientist examining how non-literal language is understood by humans and machines. Much of human communication is not encoded directly in words: we may say ‘It’s getting late’ to politely indicate that we would like to leave, or call a ballerina a ‘swan’ to capture her grace. This kind of language is called pragmatics and often thought to comprise everything from humor, to white lies, metaphors, implicature, prosody, and more. Although there have been investigations into each of these phenomena individually, it is unknown whether they are supported by the same mechanisms, nor how they relate within the individual, which has implications for neurodivergent individuals facing challenges in communication. And, while current language models show impressive results, little research has explored what kinds of computations might support pragmatic language understanding, which is crucial for success in artificial intelligence. Across both behavioral and computational approaches, the current project will address these limitations by identifying clusters of related pragmatic inferences in humans and exploring how they are computed in models. This project applies new methods to shed light on a unified framework for pragmatic language, with the goal of including neurodivergent individuals as researchers in the process. Large-scale individual-differences studies will develop a comprehensive battery of pragmatic language and expose relationships between abilities (e.g., does understanding white lies correlate best with understanding of irony?). Non-linguistic cognitive assessments will identify which clusters relate to social reasoning, literal language understanding, and executive abilities. By evaluating current language models on non-literal language interpretation, the project with also uncover the learnability, scope, and generalization of language models’ performance on pragmatic language, and will allow us to compare model activity to the clusters found in the human studies. By testing pragmatic language in humans and current computational models, this research will bring us closer to understanding the necessary input and computations for pragmatic language comprehension.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.