Keeping Colliders in Physics: Rebuilding the ALA taxonomic backbone

What this was about

Cameron Slatyer explained how the Atlas of Living Australia (ALA) rebuilt its taxonomic backbone. It had been assembled by an automated 'large taxon collider' that merged the ABRS National Species List with the Catalogue of Life, the New Zealand Organism Register and ad hoc sources. He argued that a backbone for matching data is different from a taxonomic checklist, and that a fully automated, human-free backbone is a fiction. ALA stopped the colliding: it now adds sources one at a time around the NSL, with metric checks and feedback loops through ChecklistBank and data matching. This sharply cut duplicates and unmatched records.

Why it matters. Names matching decides whether millions of occurrence records can be found and used, including for threatened species assessments, sensitive data handling and AI. ALA's metrics-driven approach is a practical model for other aggregators.

In the room

• ALA holds more than 184 million records. Its backbone is built on the ABRS National Species List because of endemism, Australia-specific taxonomy, undescribed phrase-name species, names fixed in legislation, invasive species risk and traditional ecological knowledge.
• The large taxon collider merged checklists automatically. It relied on consistent hierarchies and complete name combinations; any variance produced errors and duplicates.
• Example: records of the endangered common bent-wing bat, long misidentified as Miniopterus schreibersii, did not connect to the endemic species because the misidentification was not a synonym in the NSL.
• Listing a misapplied name both under the taxon it was misapplied to and in its correct place, which is common in checklists, is a disaster for data matching.
• The old backbone had jumbled concepts and duplicates. More than 10 million records (10% of holdings) matched nothing, and one name had 78 variants.
• A backbone needs a single hierarchy, not necessarily a fully scientifically valid one. It needs names inserted (prioritising legislated species), concepts removed or moved to avoid mis-routing records, and sometimes a commonly accepted name used over the best current name.
• ChecklistBank metrics (duplicates, orphan taxa) plus data matching (higher matches and no matches) guide curation. The team extracted all 2.06 million names-match parameter combinations.

Notable moments