Baselines

Task 1 — global alignment

Partiality-corrected P/R/F1

Headline scores are computed against the repaired silver standard R⁺_≈; scores against the unrepaired R_≈ are secondary. Metrics score all named-entity equivalences in the reference — classes, properties, and individuals alike — see evaluation metrics. Eight matchers cover all six pairs; five further matchers were run only on the two openly redistributable JC3IEDM pairs and are listed separately (their numbers are not comparable to the full set).

Micro-F1 on R⁺_≈ — all six pairs

Comparable set — 8 matchers, all 6 pairs

Headline = micro-F1 on the repaired reference R⁺_≈; the unrepaired R_≈ columns are secondary. Columns sort on click.

Partial coverage — 5 matchers, JC3IEDM pairs only (n=2)

Matcher	n pairs	Micro-P (R⁺≈)	Micro-R (R⁺≈)	Micro-F1 (R⁺≈)	Macro-F1 (R⁺≈)
LogMapLLM	2	0.996	0.877	0.933	0.946
ALOD2Vec	2	0.963	0.255	0.403	0.592
ATMatcher	2	0.926	0.242	0.384	0.537
Fine-TOM	2	0.625	0.161	0.256	0.263
KGMatcher	2	1.000	0.084	0.155	0.268

Run only on jc3iedm-mio and jc3iedm-facility; not comparable with the six-pair set above. Micro pools TP/judgeable/|R| across the two pairs; macro is the unweighted mean of the per-pair F1 values.

Best matcher per pair

“Coherent (TBox)” is the TBox-level coherence of the merged ontology pair under the system's alignment (complete reasoner, individuals removed; data: coherence_best_per_pair.json); with-ABox effects are reported separately in the supplementary analysis. Full matcher × pair detail: global_detail.tsv · global_aggregate.tsv · global_results.json.

Reference sizes

Unique equivalence pairs per reference. “Repaired: —” means the silver was already coherent at the TBox level (R_≈ ≡ R⁺_≈); see how the repaired reference was constructed.

Reported analysis: TBox-only scoring ablation

Because 60% of the headline reference consists of individual–individual correspondences, we also publish a counterfactual in which every individual cell is removed from both the alignments and the references and everything is rescored. The reference shrinks (|R⁺_≈| 488 → 190) and the top ranks reorder — the gap between the two views is itself a finding about instance matching. Read the TBox-only ablation →

Task 2 — local equivalence ranking

Hits@k and MRR

Reference baselines with a fixed abstention threshold τ=0.5; references are private. The headline is the unweighted macro-average over the two pairs. The silver is lexically clean (same-label equivalences dominate), so the naive string baseline beats the neural encoder — the contribution is the task design.

Macro-averaged over both pairs

Per pair

Matched-only / NIL-only breakdown

Pair	Baseline	Subset	Queries	Hits@1	MRR
uco-stix	lexical	matched	194	0.954	0.977
uco-stix	lexical	NIL	23	0.609	0.719
uco-stix	MiniLM	matched	194	0.789	0.873
uco-stix	MiniLM	NIL	23	0.217	0.465
stix-d3fend	lexical	matched	38	0.816	0.890
stix-d3fend	lexical	NIL	35	0.686	0.805
stix-d3fend	MiniLM	matched	38	0.789	0.871
stix-d3fend	MiniLM	NIL	35	0.400	0.605

Metrics are reported overall plus matched-only, NIL-only, and per-type; see evaluation metrics and the Task 2 description.

Editions

Baselines and results by edition

DISO-OAEI 2026 is the first edition, so there are no previous years yet. As the track matures, each edition's final baselines and participant results will be archived here, and changes to the datasets and results between editions are recorded in the changelog.