Publication
Grounded Label Space Engineering for Knowledge-Centric Annotation Workflows
Pratik Sitapara; Thiago Gouvêa; Daniel Sonntag
In: Proceedings of the 10th International Workshop on Annotation of Real World Data for Artificial Intelligent Systems. International Workshop on Annotation of Real World Data for Artificial Intelligent Systems (ARDUOUS-2026), located at KI-2026, August 11, Bremen, Germany, Communications in Computer Science and Information Science (CCIS), Vol. 7899, Springer, 2026.
Abstract
Annotation is a central bottleneck in data-centric AI, especially in expert-driven domains where labels encode evolving domain knowledge rather than merely task-specific class names. This challenge is particularly visible in domains such as bioacoustic monitoring, where experts annotate species vocalisations, environmental sounds, uncertain events, and locally specific phenomena under changing ecological and taxonomic conditions. However, for such expert-driven domains, most annotation workflows still rely on label spaces that are flat, fixed, and weakly grounded in shared semantic resources. Such label spaces support operational efficiency, but they provide limited support for representing semantic structure, linking labels to domain knowledge, or adapting to novel and ambiguous observations. In this challenge paper, we argue that current workflows treat label design as a preliminary schema-definition step, while the annotation process itself is where experts often encounter novelty, refine distinctions, and negotiate meaning. This separation causes epistemic loss: newly observed concepts, boundary cases, and semantic justifications are either forced into inadequate labels or left outside machine-actionable representations. We propose Grounded Label Space Engineering (GLSE) as a problem formulation and methodological direction for coupling annotation with knowledge representation. GLSE treats label spaces as ontology-linked semantic objects that can be elicited, grounded, revised, and stabilised under expert supervision. Rather than replacing existing annotation practices, GLSE couples annotation and knowledge engineering through controlled co-evolution, thereby addressing the tension between expressive flexibility and semantic rigour. We position this as a foundation for adaptive, interoperable, and knowledge-centric AI systems.