The TBox-only scoring ablation

Key numbers

Task 1 — comparable set

Rank movement (8 matchers, all 6 pairs)

Micro-F1 against R⁺_≈. The LogMap family's published lead rests substantially on matching the jc3iedm-mio individuals (91% of LogMap's true positives there), while every other matcher's recall was capped by a denominator dominated by cells they never attempt.

Full aggregate comparison

Micro-averaged against R⁺_≈ (the headline). The two coverage sets stay separate, as on the baselines page.

The filter

Reference shrinkage

Cells are removed when either entity is a pure owl:NamedIndividual (punned entities — also declared as classes — count as terminological and are kept). The last column is the maximum recall available today, on the published references, to a system that emits no individual mappings.

Best matcher per pair

Interpretation

Why both views are published

Both boards are legitimate — they answer different questions. The published semantics score the references as they are, in which the JC3IEDM pairs are substantially an instance-matching test: the individual cells are quasi-terminological content (country and category-code enumeration individuals, QUDT/OWL-Time units) that these ontologies happen to model as owl:NamedIndividual. The TBox-only view shows what remains once that dimension is removed: schema-matching quality the published board partly masks — the BERT-family systems' apparently poor recall hides competent schema matching, and the podium reorders on four of six pairs. Participants should read the task documentation carefully: individuals are scored, and a system that strips assertional mappings forfeits up to 86% of a pair's recall (see the table above).

Provenance

Computed by filtering pure-individual cells from both the matcher alignments and the references, then rescoring on the production metric's unchanged code path. Entity typing was taken from the published OWL files (pure individual = NamedIndividual ∧ ¬Class ∧ ¬ObjectProperty ∧ ¬DatatypeProperty; e.g. Brick: 6,363 pure, 2,193 punned kept). Verified independently: all 58 matcher×pair rows were reproduced from the original alignment files with a separate set-based implementation, agreeing to 10⁻⁹, and the unfiltered control reproduces the published baselines exactly. The full result files (per-pair detail, aggregates, and the filter log) are part of the supplementary materials (available at track launch); the published baselines are on the baselines page.