The success of modern AI systems has generated a lot of excitement about their transformative potential for science. Two major flavors of excitement include: (1) AI systems as sources of new insights about humans; and (2) AI systems as tools to accelerate the process of conducting science. This talk will introduce two projects corresponding to these two directions of excitement, specifically in the field of language sciences (Linguistics and Natural Language Processing, respectively). Through the first project, I will argue that in order for model-based insights to be useful beyond mere associative inferences, appropriate bridges must be built to connect the cognitive sciences of humans and machines. And one such bridge is developing hypothesis generation frameworks that are sufficiently concrete to support translation to human experiments. Through the second project, I will argue that while current AI systems show great potential in accelerating the implementations of in-silico experiments, we must be cautious with full automation especially for research questions whose conclusions should be drawn based on controlled experimental manipulations (as opposed to metric optimization-oriented research questions). Taken together, these projects call for more stringent frameworks for the transformative potentials of AI for science to materialize.
Najoung Kim is an Assistant Professor at the Department of Linguistics and an affiliate faculty in the Department of Computer Science at Boston University, where she leads tinlab. She received her PhD in Cognitive Science at Johns Hopkins University. She studies meaning in both human and machine learners, and is interested in understanding the nature of computation and representation that underlie minds. Her research has been supported by NSF, Google, and Open Philanthropy.