The reference & how it is repaired

The global task (Track 1 · Subtrack 1) scores your alignment against a gold reference. Because that reference is derived from UMLS and Mondo, it can be logically incoherent — so we repair it. This page explains the idea; the evaluation metrics page has the formal definitions.

Why a gold reference needs repairing

The reference equivalence mappings are assembled from the UMLS Metathesaurus and Mondo. Each source is reasonable on its own, but merging them can create logical incoherence: taken together with the two ontologies' own axioms, some reference mappings imply that a class must be both something and its own disjoint opposite — an unsatisfiable class. A system that reproduced such an "incoherent" reference exactly would be rewarded for producing a logically broken alignment. So before scoring, the reference is repaired into a coherent one.

The repair: a union of what three repair tools remove

Rather than trust any single repair tool, we take the union of the mappings removed by three independent, well-established systems run over the reference:

  • ALCOMO (under the ELK reasoner) — a dedicated alignment-debugging tool;
  • LogMap (its repair/DEBUG mode) — which can also weaken a mapping's relation rather than delete it;
  • AML (AgreementMakerLight) — a structural repair.

A mapping is treated as problematic if any of the three removed it. This union is deliberately conservative: it errs toward flagging a mapping rather than silently keeping an incoherent one.

Option-Two annotation: keep, weaken, or set aside

We do not simply delete flagged mappings — that would throw away information. Instead, each reference mapping is annotated (“Option-Two”):

  • Survivors keep their relation. A mapping no tool removed stays as =; where LogMap weakened a mapping, the reference keeps that weaker < / > (subsumption takes priority over deletion).
  • Fully-removed mappings become ? (“unknown”). They are neither correct nor incorrect — they are simply set aside during scoring.
  • owl:deprecated classes are dropped from both the reference and predictions as out-of-task.

What this means for scoring

Systems are scored against both references:

  • Against the standard reference — traditional P/R/F1 (every mapping counts).
  • Against the repaired reference — coherence-aware and relation-agnostic: ? mappings are ignored from both the precision and recall denominators, and a reference </> is credited by a predicted correspondence of any relation.

The repaired, coherence-aware F1 is the headline for the global leaderboard; the standard F1 and a reasoner-checked Global Coherence score (does your own alignment leave the merged ontology satisfiable?) are reported alongside. See the evaluation metrics and the baselines for the numbers.

The repaired references are built by a consensus of the three tools per pair. Global Coherence is computed organiser-side (it needs an OWL reasoner, which does not run in the lightweight submission container), so the coherence column is filled in after the offline pass.