Type: article
Abstract: Labeled data operationalize how real-world phenomena are represented in computational systems and underpin artificial intelligence (AI) models. Prior research largely adopts a consensus-oriented view of ground truth, overlooking divergent interpretations in domains that are inherently uncertain and contested. Drawing from representation theory, we examine ground truth uncertainty and develop a framework explaining how data label divergence arises from two forms of representational inadequacy: schema inadequacy and evidence inadequacy. We further theorize that the effects of representational inadequacy are contingent on annotators’ domain knowledge. Using a controlled experiment, we find that schema inadequacy increases divergence, whereas evidence inadequacy reduces divergence. More importantly, domain knowledge plays a contingent role: when schema and evidence are aligned (both adequate or both inadequate), domain knowledge reduces divergence; when misaligned, domain knowledge amplifies divergence. These findings show that divergence is a systematic outcome of representation and interpretation, advancing research on data work and data labeling.
Source: AIS Electronic Library (AISeL)
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7222989808
