NSF Award
2213758
Humans have different kinds of knowledge about the languages they speak. Phonotactic knowledge refers to their knowledge about sound patterns in words. Such knowledge is usually studied through behavioral experiments such as wordlikeness judgment tasks, in which a speaker is asked to indicate whether a nonce (made-up) word form could be a real word in their language. Such studies have observed that wordlikeness judgements are gradient; that is, they form a cline of acceptability. This project examines different types of computational models and experimental designs to better understand the nature of the grammatical system that underly this variety of gradient judgment. For example, one class of computational models seek to account for this gradience in wordlikeness tasks by incorporating fine-grained generalizations in sophisticated statistical models. In contrast, other models based solely on categorical generalizations appear to do just as well in accounting for behavioral wordlikeness findings. This suggests that gradience observed in the behavioral experiments may reflect the experimental setting rather than the underlying grammar of the knowledge itself. This collaborative research project develops a thorough investigation of these contrasting accounts of gradience in human wordlikeness judgements. The project provides mentoring and research training opportunities for graduate students in modern experimental and computational methods; and it creates an open-access, online database of the project's experimental results that serves as a resource for other researchers in the language sciences. In addition, the investigators plan to engage in outreach activities targeting members of the public in settings focusing on informal science education.<br/><br/>This research project deconstructs the variety of factors involved in wordlikeness judgements. It includes experimental studies that manipulate response measures and stimulus and response modality to elicit human-subjects' wordlikeness judgments for nonce forms from speakers of English and Korean. A variety of computational models are assessed by comparing how well they can be fitted to the results of the experimental studies. The project's interconnected studies aim to answer the following questions: (a) how do different explicit computational models perform against observed wordlikeness judgments, (b) to what degree is gradience in acceptability judgments related to the probabilistic nature of human perception, and (c) are the results obtained affected by the manner in which words forms are presented to speakers (e.g., auditorily vs. visually) and/or in which speakers record their judgments?<br/><br/>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.
