Abstract: Researchers and practitioners are increasingly building and deploying large scale systems that are making determinations on ill-defined and complex concepts at scale either implicitly or explicitly. Is it ethical or moral for my generative model to answer this user query? Is this community post/comment toxic or harmful? To make this possible, we often rely on asking groups of people to judge and annotate data that is then used for training or evaluation. However, many traditional tools and workflows often assume a clear, unambiguous "ground truth" answer and can fail to accurately capture group judgments on subjective, ambiguous, and socially-situated concepts. Instead of ignoring disagreement or choosing arbitrary definitions/guidelines, processes for alignment tasks over fuzzy, socially-constructed concepts need to be able to capture misaligned interpretations and then resolve them through social processes for reconciliation. In this talk, Chen will present research on enabling collective alignment of socially-constructed concepts at scale, including novel annotation tools that capture sources of uncertainty separately---paving the way for targeted interventions to resolution, as well as a novel approach to social alignment that complements current constitutional approaches by grounding decisions over a body of precedents rather than abstract rules.
Bio: Quan Ze (Jim) Chen, Ph.D. is currently a postdoc in the Allen School (CSE) at the University of Washington. His dissertation research centers around building tools and workflows to understand and address various sources of uncertainty in subjective and complex human judgments. He is also broadly interested in creating human-in-the-loop AI systems that can better reflect the values of groups and communities when working with subjective and socially-situated tasks such as online moderation and misinformation response. Jim also completed his Ph.D. at UW advised by Amy X. Zhang. Jim is on the job market this cycle seeking academic or industry opportunities.