Task Decomposition for Efficient Annotation

Decomposing structured annotation tasks across heterogeneous annotator pools (models plus humans with varying expertise) reduces per-annotator inferential load and improves label quality without proportionally increasing cost. The paper formalizes how to split complex annotation into subtasks allocated by annotator capability, targeting large-corpus structured representation labeling where full end-to-end human annotation is prohibitively expensive.