DARS: An Approach for Data Annotation Requirements Specification

Yi Peng; Hina Saeeda; Hans-Martin Heyn; Jennifer Horkoff; Eric Knauss; Fredrik Warg

doi:10.1109/REW72749.2026.00020

DARS: An Approach for Data Annotation Requirements Specification
Paper i proceeding, 2026

Data annotation is a critical, yet often overlooked process in the engineering lifecycle of AI-enabled software systems. Recent interviews with industry experts revealed that the data annotation process introduces distinct requirementsrelated challenges regarding ambiguity, context specification, and verification. Here, we propose the Data Annotation Requirements Specification (DARS) framework with two artifacts: (1) the Annotation Negotiation Card to align diverse stakeholders on system objectives and constraints; and (2) the Scenario-based Annotation Specification, a template for atomic, verifiable data annotation requirements. We evaluate DARS through an empirical study involving twelve participants with AI expertise from both industry and academia. Our results demonstrate that DARS mitigates ambiguity and identifies blind spots overlooked in annotation workflows by providing a structured link between system goals and data labeling. This suggests a strong practical relevance for industrial practice.

Data Annotation Requirements Specification (DARS)

Data annotation

Annotation Negotiation Card