Medical Subject Headings (MeSH)

Controlled descriptors and hierarchy used to index PubMed records.

Rows
350K
Scope keyword
mesh

Ask Academica for a report and name this source to scope the search to it.

Live Index Search (mesh)

Three models, with one added evidence layer each

Every version uses the same training split and candidate pool. Only the ranking evidence changes.

Model 1 · Control

Frequency formula only

A fixed equation adds token-to-MeSH association scores.

  • Token-MeSH pair frequency
  • Association lift
  • Phrase-length weight
  • Cross-source bonus
  • No learned ranking
  • No hierarchy evidence
Purpose: established baseline
Model 3 · Geometry test

GBM + hierarchy + Poincaré

The GBM also receives compact coordinates learned from the full MeSH graph.

  • All Model 2 features
  • Candidate Poincaré coordinates
  • Token-associated centroid distance
  • Nearest candidate distance
  • Radial depth signal
  • Cross-branch geometric separation
Purpose: prove whether vectors add value

Why a hierarchy needs different space

A hierarchy grows exponentially. Every parent can create several children, which create more descendants.

Euclidean space

Equal physical distance represents equal vector distance. Deep branches need increasing area and become crowded.

Poincaré disk

Available space grows toward the boundary. The disk can separate many descendants while keeping parents nearby.