Methods & FAQ

How the atlases are built

Each atlas pairs a fixed census of taxa or a theme with the indexed literature linked to it, read by an expert-model panel against a published codebook and audited. This page answers the technical questions: where the data comes from, how it is linked and tagged, how accurate it is, what the counts do and do not mean, and how to cite and reproduce it.

What is an atlas?

An atlas is a map of research effort over a defined subject. It starts from a census — every accepted species of a plant family, every genus of a fossil group, or a thematic corpus such as land degradation — and links each member to the indexed papers about it. Every paper is then tagged against a controlled codebook, so the corpus can be queried by question, not just by keyword.

Where do the papers come from?

Bibliographic metadata (identifiers, titles, years, work types, authorship, citation counts) comes from OpenAlex, an open index of scholarly works, released under CC0. The record is the indexed, largely English-language literature. It is not all research: work that is not indexed, not in English, or published outside the indexed journal system is under-represented. A count of papers is a measure of indexed research effort, not of knowledge.

How are papers linked to the census?

Papers are linked to a taxon or theme by matching names in the title and abstract, guarded against false matches (a one-word genus name that is also a common word, for example), and each candidate link is judged for its role: is the taxon the subject of the study, merely mentioned, or unrelated. Only subject and substantive-mention links are kept. Synonyms are resolved to the accepted name so a species' literature is not split across its old names.

How are the tags assigned, and how accurate are they?

Each paper is read by a panel of domain-expert models against an expert-authored codebook of controlled dimensions (the subject, methods, threats, governance, and more). A proposed tag persists only if it is in the codebook, grounded in an evidence span in the title or abstract, and confirmed by a second model of a different family acting as an adversarial judge. Accuracy is measured by blind precision: a third independent model, which never acted as tagger or judge, re-judges a random sample. Each atlas states its own measured precision and interval in its methods note. The measurement is taken on the abstract-bearing corpus, because the judge needs an abstract; title-only papers are tagged but sit outside the instrument, and each page says how many.

What do the counts mean — and not mean?

The counts describe how much indexed research exists on each subject, not how much is known and not whether a subject is neglected. Three systematic confounds pull the numbers around: English-language indexing favours some regions and literatures; naming and taxonomic lag means recently described taxa have had less time to accrue papers; and, for the fossil record, the rock record itself limits what can be studied. Where an atlas discusses gaps, it states these confounds rather than reading a low count as neglect.

How are study sites located?

Study sites are extracted from the abstract — where the work was done, not where its authors are based — and resolved against open gazetteers (GeoNames, OpenStreetMap). Each page states how many of its studies name a locatable site; many do not, so the mapped subset is smaller than the corpus and is labelled as such.

How current is an atlas?

The atlases are living resources. Each carries a “last updated” date under its header, and every figure is re-derived from the database and audited before it changes. The census of taxa is refreshed separately, when the source authority releases a new version. Because the atlas keeps changing, citations should carry the version and the access date.

How do I cite an atlas, and is the data reproducible?

Each atlas carries a recommended citation and a BibTeX entry at the foot of its page. Where an atlas has been archived, it also has a persistent identifier (DOI) with two forms: a concept DOI that always resolves to the latest archived version, and a version DOI for the exact release you used. The DOI resolves to a fixed, downloadable snapshot on Zenodo — the census, the linked papers and their judge-confirmed tags, the study sites, and the figures behind the page — so a reader can reproduce the atlas as it stood at that version. The live atlas keeps updating; the archived snapshot does not, which is what makes it citable.

What licence applies?

COMPASS ID's own output — the tags, the linkage, the figures, and the atlas page — is released under CC-BY-4.0. Bibliographic metadata from OpenAlex is CC0 and is credited; abstract text belongs to the publishers and is not redistributed. The taxonomic census is credited to its source (for plants, the Kew World Checklist of Vascular Plants), and no restricted source tables are redistributed. Study-site coordinates derive from OpenStreetMap (ODbL) and GeoNames (CC-BY); archived snapshots carry site names and countries, with coordinates on the live atlas.

What are the main limitations?

The record is the indexed, largely English-language literature, so absence of papers is absence of indexed evidence, not proof that a subject is unstudied. Automated tagging is high-precision but not perfect, and its accuracy is measured on abstract-bearing papers only. Links rest on names in the title and abstract, so a study that discusses a taxon without naming it in either is missed. Geography is limited to studies that name a locatable site. Every atlas states its own figures, intervals and caveats in its methods note; read it alongside the maps.