Library and information science

Why this page leads the set

Library and information science is not a vertical. It is the discipline the other nine pages borrow from, which is why it is the first entry rather than a peer of the rest.

The case for that position is not sentimental. In the talk cited below, Jessica Talisman puts it at 3:22: as the 1990s big-data movement optimised for storage first and query layers second, library and information science was already building metadata ecosystems threaded through machine-readable records for governance, reliability and consistency. Libraries went online in the 1960s, before the web, because worldwide catalogues had to be searchable and their records reconcilable. Those systems are still in service, connecting libraries for lending and access.

The practical consequence for an agentic programme is a sourcing rule. When a team starts modelling a domain and reaches for a bespoke vocabulary, it is usually about to re-solve a problem with a published rulebook. The vocabulary stage of the pipeline is a library-science stage, and the rulebook it cites is a library-science standard.

Where the method comes from

At 9:04 the pipeline's first stage is handed a construction standard rather than a modelling preference: a guideline for the construction, format and management of monolingual controlled vocabularies, to be read alongside the W3C's SKOS documentation. Stated aims, in the same passage: eliminate ambiguity, control synonyms, test and validate, and establish relationships between concepts.

That is ANSI/NISO Z39.19, and it is the document that turns a glossary into a controlled vocabulary. Three references carry the stage:

• ANSI/NISO Z39.19 — the construction rulebook: preferred terms, aliases, definitions, scope notes, and the discipline of deciding what is one concept and what is two.
• ISO 25964-1 and -2 — the thesaurus standard, and the standard for interoperability and mapping between vocabularies. Part 2 is the one to read if the plan involves aligning a local vocabulary to a published one.
• W3C SKOS — the encoding that makes a vocabulary machine-readable, with preferred, alternative and hidden labels, definitions, and broader, narrower and related relations.

Two of these already have pages on this site, at W3C SKOS and ISO 25964. This page covers what the rest of the domain set does with them.

What is reusable, by layer

Read as a whole, the table is an argument. Every layer in it corresponds to something an agent now needs, and almost every one of them predates the agent.

The playbook

Reuse

Align to ANSI/NISO Z39.19 for construction, ISO 25964 for thesaurus structure and mapping, and SKOS for encoding. Do not write your own vocabulary rules, and do not reach for OWL before a concept scheme has been tested.

Scope

One bounded term space with a visible boundary: one product catalogue, one internal function, one reporting vocabulary. Not the organization's vocabulary.

Competency questions

Is this one concept or two? Which label is preferred and which are aliases? Who owns this definition? Which published vocabulary should this local term map to?

Smallest artifact

A concept scheme: preferred label, alternative labels, definition, scope note, owner, source. Publish it as SKOS, and keep it in version control.

Validation

Have the domain experts adjudicate a sample of mappings and definitions, and record their disagreements. The disagreements are the requirements; the agreements are usually already known.

Characteristic failure

Synonym collapse and silent duplication — two terms that mean the same thing, or one term that means two. Neither raises an error, and both degrade retrieval for as long as they persist.

Limits of this page

• The domain is standards-rich, which makes reuse cheap, and adoption-poor inside AI teams, which makes the work unfamiliar. Standing behind a vocabulary nobody curates is worse than having none, because it implies a check that is not happening.
• The pipeline described below stops at the semantic structures. Library science does not supply validation, permission, currentness or agent coupling; those have to come from elsewhere in this set.
• Several of the sources are cited by name only, because their publishers do not serve the pages to automated clients. The links that are published here were fetched and their bodies inspected.