Keet notes, chapter 7: bottom-up sources need lifting and review

Chapter 7 of Maria Keet’s ontology-engineering textbook covers bottom-up development: building ontology from what already exists, whether that is a relational database, a document store, a spreadsheet, or the output of a mining tool. Our study passes through the chapter kept meeting one line worth keeping: bottom-up work is semantic lifting, not schema export. A live system expresses the business imperfectly, through denormalized tables, overloaded status fields, foreign keys that only hint at relationships, and legacy compromises. A direct conversion turns all of that into something formal-looking while formalizing storage habits instead of domain meaning. Someone still has to recover which structures reflect entities, which values are really controlled vocabularies, which rows should stay instance data, and which source assumptions need to stay visible after conversion.

Three operating rules fell out of the lifting sections. Preserve provenance: a generated OWL artifact can look final even when it is one projection of a source, so the source, transformation path, and known compromises stay on record. Make class-vs-individual decisions explicit: a status column with lead, quoted, won, lost should not become ontology classes just because the values recur in many rows; the column may be a controlled vocabulary worth modeling, a collapse of several business decisions, or an application shortcut the company wants to replace, and if the platform does not record the decision, the choice still gets made, just invisibly. Keep review surfaces label-aware: domain experts cannot review a converted asset that shows raw identifiers instead of meaningful labels.

AI proposes; review decides

The chapter’s semi-automated methods, text mining, clustering, document parsing, reasoning services, and LLMs, can surface candidate terms and relations a team would take much longer to see. That is real value, and it comes with the chapter’s sharpest governance line: the machine may propose the category, but a reviewed business process decides whether the category becomes shared memory.

The illustrative case we used in the original study note (an invented scenario, not an observed client): an AI tool reads a clinic’s intake notes, explanations of benefits, and claim denials, and proposes one clean category called Patient Follow-Up Failure. The label is countable, routable, and assignable, and it groups a missed form, a payer-specific attachment requirement, a plan rule the clinic never captured, and a clinical-signoff lag. Four operating problems with four different owners, erased by one plausible name. If that label goes straight into the ontology, the next operator inherits a false business memory: reports count the wrong thing and fixes aim at the wrong owner.

So the rule is queue, not authorship. A candidate record carries its own trail: the source document that triggered the suggestion, the exact term or relation proposed, the nearby evidence, the model or heuristic that produced it, the reviewer, and the decision, accepted, revised, deferred, or rejected. The trail is not bureaucracy; it is how domain expertise stays in charge while automation helps with scale, and how a later operator can inspect why a category was split without interviewing the person who split it.

The gate sits before promotion

Semi-structured sources make the sequencing question concrete. Document stores look closer to real-world nuance than rigid tables, property graphs look like ready-made domain structure, and spreadsheets are approachable enough to hand to any operator. That flexibility lets several problems hide at once: the same idea under different labels, optional fields implying unstable cardinality, conflicting assertions with no repair path.

The lesson our 2026-04-12 pass turned into ONT-026 (an opened governance issue, still planned work, and it may change shape before it ships): contradiction review, duplicate handling, and cardinality review have to happen while the assertion is still a candidate. If a source is lifted into ontology form first, the platform is already treating a source quirk as business structure, and downstream validation ends up checking a commitment that should still have been under review. SHACL-style conformance says whether data fits declared structure; it cannot say whether the structure should have been declared.

Spreadsheets get the clearest version of the rule: an authoring contract that says which columns are candidate terms, which rows may propose structured assertions, which values stay raw data, and which review step approves promotion, with a trail of what was accepted, merged, revised, or left out. When one branch manager writes Customer Tier, another writes Account Class, and a third leaves the field blank for cash customers, those columns may point at one shared concept, three local habits, or an unclear business rule. The gate is where that gets decided, rather than hardening by ingestion.

If you run a business

The ownership translation is direct. An owner-operator should inherit a business model that reflects how the company works, not a vendor’s export format, a team’s naming habits, or a spreadsheet built under deadline pressure. Skip the lifting and the review, and reusable infrastructure gets built on implementation noise: apprentices inherit concepts nobody made explicit, and later automation reasons over yesterday’s storage shortcuts. Chapter 7’s rule set, condensed: lift meaning, not schemas; let AI propose and reviewers decide; keep the candidate trail; put the gate before promotion; give spreadsheets a contract.

---

Source evidence used in this note: Maria Keet’s ontology-engineering textbook, Chapter 7 (bottom-up ontology development); smb-ontology-platform/docs/plans/2026-03-31-keet-ontology-engineering-progress-tracker.md (2026-04-03 and 2026-04-12 entries); and smb-ontology-platform/docs/issues/ONT-026-add-semantic-lifting-governance-for-source-schema-classification-and-app-boundaries.md. ONT-026 remains open, unshipped planned work that may change shape or may not ship at all. The clinic follow-up-failure scenario is illustrative, not an observed client incident. This chapter note consolidates three earlier study posts on semantic lifting, AI-proposed ontology candidates, and semi-structured source review gates.