What it is
OmniTaxonomy is a planned open-source tool for libraries and archives. Catalogs usually describe things using a few Western classification systems, which can erase the words communities use for their own concepts. With OmniTaxonomy, you enter a term in any language or community vocabulary and get back possible matches across many classification systems, each with a confidence score. It never replaces the original term. Matches it isn't sure about go to a person for review. It also aims to record concepts that have no Western match at all. The work is paused.
How it works
- Submit a term, with its language or community source.
- The system searches across classification systems in five stages: exact and near-spelling matches, matching by meaning across languages, the structure of each vocabulary, an AI re-ranking, and a confidence check.
- You see candidate headings with scores; the original term stays attached.
- Low-confidence matches are sent to human reviewers.
What exists today
- A small working converter that turns a common vocabulary format into one an alignment library can read, with thorough tests
- Decision documents on architecture, data model, consent, licensing and tech stack
- Research plans for benchmarking, AI re-ranking, confidence thresholds and a "no Western equivalent" dataset
- A primer on community data rights and an outline of a data-ethics statement
What isn't there yet
- The matching pipeline, API and data loading
- Any experiments; the research plans are not started
- Planned plug-ins for archive platforms such as Omeka S, Mukurtu and WordPress
- Planned research outputs: a tuned model, an alignment dataset and a benchmark
Under the hood
- Planned as a Python web service with Postgres, a graph database for vocabularies and a vector search for meaning-based matches
- Vocabularies in SKOS, a standard format for subject lists
- Target systems include Library of Congress headings, Dewey, UDC, AGROVOC, Getty vocabularies and Māori Subject Headings
- What exists runs as a small Python project with automated tests