Choosing between synthetic and expert taxonomies for downstream use

Ascertain which type of taxonomy—Simula‑generated (language‑model‑generated) taxonomies or expert (human‑authored) taxonomies—is better suited for specific downstream applications.

Background

The authors evaluate language‑model‑generated taxonomies against expert references using completeness, soundness, novelty, and coverage. While the comparative results suggest strong coverage and novelty for model‑generated taxonomies, the practical utility for downstream tasks is not established. The authors explicitly state that it is not clear which taxonomy type is preferable for downstream applications.

References

Downstream Application. While we performed a comparative evaluation of synthetic and real taxonomies, it is not directly clear which are better suited for certain downstream applications.

— Reasoning-Driven Synthetic Data Generation and Evaluation  (2603.29791 - Davidson et al., 31 Mar 2026) in Appendix A: Reasoning-driven Taxonomy Generation and Evaluation, Limitations (Downstream Application)