On the diverse language experiences of humans and machines

Abstract

Human experience is characterized by remarkable linguistic and sociocultural diversity. At the same time, much of this diversity is neglected in the language-related fields of linguistics, cognitive science, and natural language processing. This dissertation thus explores how diverse language experiences can lead to overlooked variation in humans and machines. Beginning with English-speaking families in the Pacific Northwest, Chapter 1 observes how children’s language environments—in particular, differences in maternal and paternal input—can lead to variation in infant volubility, illustrating the delicate relationship between experience and behavior. Chapter 2 then devises an iconic approach to artificial language learning to investigate variation in cognitive biases across diverse language communities; importantly, seeking this variation directly can illuminate aspects of cognition that may be rooted in language experience rather than innate constraints. Finally, turning to machines, Chapter 3 shows how multilingual language models can encode many morphosyntactic properties crosslinguistically, but only occasionally uncover when a property is shared by multiple languages—beckoning the questions: To what extent does this variable behavior arise from crosslinguistic variation present in multilingual training data? And to what extent is it indeed humanlike? Together, these studies underscore the importance of minding diverse language experiences for understanding language processing in humans and machines.