Authority files hold canonical names for entities and resolve homographs and aliases

What it is

An authority file gives real-world entities—people, organizations, places, works—one canonical name, plus aliases and easily confused namesakes. It repairs two accidents: the same string, different referents (two people named Zhang Wei) and different strings, same referent (Peking University / PKU / Beida). A controlled vocabulary governs words for concepts; an authority file governs identity of instances. Without an identity layer, author = Li Na stacks a tennis champion and a singer into one person, and both filters and relatedness lie.

The canonical name is the entity’s preferred written form, not its most common nickname in the press. Nicknames are aliases, not the identity key.

Why it happens

Equal strings do not entail equal objects; unequal strings do not entail unequal objects. An authority file puts a stable id under the strings, and every spelling points at the id. Homographs split on added attributes (birth year, role, works); aliases merge. Retrieval and navigation show the canonical name; the query side still absorbs aliases. People see a distinguishable entity and type the string they already know.

Without an authority layer, deduping can only use string rules, which will over-merge on namesakes and over-split on aliases. If a matrix or facet uses “author” as an axis, the values are a pot of undisambiguated strings and the cell has no interpretable entity.

Studying it

Take known namesake pairs and alias groups as gold, and see whether the authority layer splits and merges.

• Paradigms: homograph tasks (find the right author) and alias tasks (reach the same organization under different spellings); compare no authority, string normalization only, and an authority id. Library name authorities are a method source; the whole ILS need not be copied into the product.
• Independent variables: presence of a stable id, whether disambiguating attributes are visible, alias coverage.
• Dependent variables: false-merge rate, false-split rate, recovery via a nickname, whether people can say “which namesake this is.”
• Methodological note: exact-match experiments underestimate homograph harm because lab names are often unique. Sample real frequent namesakes. The cost of a wrong merge exceeds a temporary split: a false merge pollutes everything attached to that entity; a false split can still be patched with an alias.

Where it stops holding

Ordinary topic words (“expense,” “invoice”) are not entities and should not enter an authority file, or the concept layer and the instance layer tangle. One-off objects with no durable identity (a single support chat) do not repay an authority record. User-chosen display names may coexist with the canonical name, but permission and audit must bind the id, not the display name. In an open world entities keep arriving and the authority file lags; lag should appear as an “undisambiguated” state, not as auto-merging a new string into the most similar old entity.

Applying it

• Give people, organizations, places, and works an id, a canonical name, aliases, and one disambiguating gloss; show the gloss when namesakes appear.
• Aggregate filters, matrices, and relatedness on ids, not on display strings.
• On import, align to authority first; leftovers go to undisambiguated. Do not auto-merge on string equality.
• Verify with two namesake entities and three alias sets on lookup and filters. Namesakes must not share a count; the three alias sets must hit one id. Failure either way means the authority layer is not yet wired into the structure—only an extra “standard name” column.

Related

• Within the group: G1.10.1 A controlled vocabulary gives each concept one preferred term; others point to it · G1.10.2 A thesaurus describes hierarchy and association with broader, narrower, and related terms · G1.10.4 Controlled vocabularies are staff-maintained, unlike folksonomies · G1.10.5 At retrieval, non-preferred terms map to the preferred term and yield the same results · G1.10.6 Vocabularies need versioning and change records as the domain evolves
• Adjacent: G1.05 Metadata · G3.15 Result grouping and deduplication · G1.08 Terminology consistency
• Search terms: authority file · name disambiguation · homograph