back
Back to Glossary

What is ontology versioning?

Ontology versioning is the practice of evolving a label taxonomy in a controlled way so historical data stays usable while new concepts are added. In data operations for AI, the “ontology” is the structured set of classes, attributes, and relations that define what annotators can label and how those labels connect. Synonyms you will see include taxonomy versioning and label schema evolution.

Why it matters: production AI teams rarely freeze their schema. New product lines, regulations, or edge cases drive changes. Without versioning, model training mixes incompatible labels, reduces comparability across experiments, and increases rework. Knowledge engineering literature treats versioning as a first-class problem, similar to software configuration management, with methods to compare, migrate, and align schemas over time.

A practical ontology versioning workflow includes:

  1. Version identifiers and human-readable change logs for every release.
  2. Change types such as add, split, merge, deprecate, and rename, each with examples.
  3. Forward and backward mapping rules to maintain compatibility and enable evaluation.
  4. Dataset-level provenance that records which model saw which label set and why.
  5. Impact monitoring for training stability, evaluation drift, and error clusters after a change.

Tooling considerations:

Annotation platforms should store the ontology as a managed object with history, enforce compatibility checks before job launch, and offer migration scripts that map old labels to new ones. When inter-team alignment is needed, a lightweight governance board can approve changes and publish a playbook for annotators, reviewers, and MLEs.

Example:

A retailer splits a generic “Outerwear” class into “Jacket,” “Coat,” and “Raincoat.” Backward rules map the three new classes to the older parent for legacy models; forward rules instruct annotators to use the new set from a specific date. Historical benchmarks use the backward map for comparability, while the next-generation model trains on the split classes.

‍