Ontology Alignment Map¶

A semantic correspondence artifact — instantiates Structure-Preserving Embedding Design

Links concepts in one vocabulary to their counterparts in another through anchored correspondences — equivalent, broader, narrower — each carrying a human-readable rationale.

This mechanism works at the level of meaning, not fields or coordinates. An Ontology Alignment Map pins concept-to-concept correspondences — this source concept is equivalent to, narrower than, or overlaps that host concept — and its defining feature, the thing that separates it from a schema table, is that every non-trivial link carries an explicit human-readable justification. Semantic equivalence is a judgement, and a judgement has to be something a person can read and challenge. A small set of high-confidence anchors — the correspondences no one disputes — calibrates the harder alignments around them.

Example¶

Two retailers are merging catalogs, and their product taxonomies do not line up. Company A's Notebooks has to align to Company B's Laptops; A's Ultrabooks is narrower than B's Laptops; A's Tablets overlaps B's 2-in-1s but is not the same thing. The map records each as a typed correspondence with a note — for Notebooks ≈ Laptops: "exact match; both denote portable clamshell computers; checked against a few dozen sample SKUs." The anchor set is the dozen-odd unambiguous equivalences everyone signs off on immediately (HDMI Cable = HDMI Cable); the contested concepts are then aligned relative to those anchors. The link types come from a real vocabulary — SKOS mapping relations such as exactMatch, broadMatch, and relatedMatch.^[1]

The outcome is a merged catalog where a shopper's search for "laptops" also surfaces A's notebooks, and — the part that matters six months on — where every non-obvious equivalence has a rationale a merchandiser can pull up and dispute, instead of a silent category collapse that quietly buried a distinction the business depended on.

How it works¶

Establish an anchor set first — the high-confidence, unambiguous correspondences — and calibrate everything else against it.
Type each correspondence — equivalent, broader, narrower, overlapping, or no-match — rather than flattening all of them to a bare "A relates to B."
Attach a rationale to every non-trivial link, so a domain expert can read why two concepts were joined and object if they disagree.
Leave the genuinely unmatched unmatched. A first-class "no faithful counterpart" verdict is preferred over forcing a false equivalence.

Tuning parameters¶

Match-type vocabulary — how many relation kinds are allowed (just "equivalent," or the full broader/narrower/overlap set). Richer typing captures partial matches but demands more judgement per link.
Anchor confidence threshold — how certain a correspondence must be to serve as an anchor. Stricter anchors calibrate better but leave more concepts to align by inference.
Rationale depth — a one-word tag versus a short paragraph per link. Deeper rationale is auditable but slows the work.
Coverage versus precision — align every concept (accepting weak links) or only the defensible ones. Pushing for coverage is how dubious equivalences creep in.

When it helps, and when it misleads¶

Its strength is that it makes cross-vocabulary integration contestable: a human can read why "Notebooks = Laptops" and push back, which is the only real defence against silent semantic drift when two conceptual systems are fused.

Its classic misuse is over-alignment — forcing equivalences to declare two systems "compatible," often to justify an integration or merger already decided, collapsing distinctions that actually mattered. Same-word-different-meaning "false friends" slip through whenever the rationale step is skipped. The discipline is to require a rationale for every non-anchor link, keep the "no faithful match" verdict first-class, and let a domain expert — not the integrator racing a deadline — sign off on the hard ones.

How it implements the components¶

This mechanism realizes the meaning-correspondence side of the archetype — the concept links and the human account of them:

anchor_correspondence_set — the calibrated set of high-confidence seed correspondences that pin the alignment and anchor the inferred links.
human_explanation_layer — the per-link rationale that makes each semantic equivalence readable, auditable, and open to challenge.

It does not test whether the mapped relations survive under host operations — that is the Invariant Preservation Test Suite; the concrete field-and-type table is the Schema Mapping Table; and reconstructing the original source form is Serialization with Reconstruction Schema.

Instantiates: Structure-Preserving Embedding Design — the map is the semantic layer of the embedding, where concept meanings are made to correspond.
Sibling mechanisms: Schema Mapping Table · Invariant Preservation Test Suite · Structure-Preserving Map Specification · Vector Embedding Model · Round-Trip Validation Test · Serialization with Reconstruction Schema · Adapter or Wrapper Layer · Coordinate Chart Mapping · Graph Embedding · Embedding Collision Probe · Nearest-Neighbor Audit

Notes¶

This is where the archetype's neighbour distinction bites hardest: an alignment map preserves semantic relations — a concept's meaning and its broader/narrower links — which is not the same as preserving locality in a metric space. A vector that sits nearby is not thereby a synonym, so an alignment must be justified on meaning, never inferred from proximity alone.

References¶

[1] SKOS (Simple Knowledge Organization System), a W3C standard, provides mapping properties — exactMatch, closeMatch, broadMatch, narrowMatch, relatedMatch — that give a shared, typed vocabulary for correspondences between concept schemes; using them is what turns "these two are related" into a specific, checkable claim. ↩