Acquisition
Systematic literature search across scholarly sources — OpenAlex, CORE, Semantic Scholar — with a citation backbone (OpenCitations) and citation-based ranking to build a curated, non-redundant corpus.
Plate IILiterature → Data
Nomia Library
A curated, structured repository that aggregates and standardizes dispersed taxonomic literature — transforming scientific PDFs into provenance-anchored, machine-readable knowledge.
⟢ Overview
From scattered PDFs to structured science.
Foram taxonomy is documented across thousands of journal articles, monographs, and legacy catalogs — many decades old, in multiple languages, with inconsistent terminology. Nomia is the ingestion pipeline that solves this: it reads that literature systematically, extracts the relevant biological and bibliographic content, normalizes it against a controlled vocabulary, and outputs structured records that downstream tools — Testaria, the educational portal, and the open API — can query reliably.
The pipeline is designed around a clear separation of concerns: acquisition finds the literature; extraction reads it; normalization aligns terminology; structuring stores it. Each stage has defined quality gates, and no record advances without passing the stage before it.
Status — WoRMS data layer ingested · Full pipeline in design
⟢ Pipeline Stages
Four stages from source to structure.
Extraction
PDF → Markdown conversion (GROBID/Docling-class tooling), bibliographic and scientific metadata extraction, species-name detection (GNfinder-class), and candidate morphological-character extraction from text with mandatory human review.
Normalization
Lexical variation across publications — 'foliate' vs 'foliáceo,' 'bilamellar' vs 'bilaminar' — is mapped onto a controlled SKOS vocabulary (prefLabel / altLabel) so every concept resolves to a single canonical term.
Structuring
Normalized data is written to PostgreSQL + pgvector and a faceted OpenSearch index, keyed by WoRMS AphiaID. Every record carries a provenance reference; nothing is asserted without a source.
⟢ Design Principles
Non-negotiable data rules.
These constraints are architectural, not optional. Every component of the pipeline is designed to enforce them automatically; violations surface as quality-gate failures, not silent data errors.
- Provenance
- Every datum carries a direct pointer to its source reference. Unsourced assertions are not admitted to the database.
- Morphology ≠ Interpretation
- Observable morphological characters are recorded separately from taxonomic interpretations. The model does not conflate what is seen with what it means.
- LLM Role
- Language models translate text into structured extraction candidates. They are never taxonomic authorities — a human curator confirms every record before it enters the permanent store.
- Controlled Vocabulary
- Character states are drawn from a governed SKOS vocabulary. Free-text state values are not permitted; all terms must resolve to a prefLabel.
- AphiaID Keying
- All species records are keyed by WoRMS AphiaID. Taxonomy is never forked; synonymy and hierarchy sync to the live WoRMS authority file.
- Open Access
- The structured output — bibliographic records, character occurrences, vocabulary — is fully open. Only content under restrictive third-party licences is excluded from redistribution.
"The LLM translates text into structured extraction candidates. It is never a taxonomic authority — a human curator confirms every record."
⟢ Outputs
What the pipeline produces.
Bibliographic corpus
Structured reference records with provenance links, de-duplicated and ranked by citation weight.
Character occurrences
Morphological character–state pairs extracted from text, linked to the source passage and the species AphiaID.
Controlled vocabulary
A SKOS vocabulary of character labels and state terms, with synonym (altLabel) mappings for cross-publication normalization.
⟢ Continue Exploring
See how Nomia's output powers identification.
The structured character records Nomia produces are the foundation Testaria queries when it filters species candidates by morphology.