At retrieval, non-preferred terms map to the preferred term and yield the same results

What it is

If the vocabulary only binds cataloguing, a user who types a non-preferred term still gets zero results. Retrieval mapping requires that a query hitting a non-preferred term be rewritten or expanded internally to the preferred term (and to that concept’s id), with a result set identical to submitting the preferred term. Mapping happens at retrieval; it need not change the query string the user sees. The visible words may stay; the hit concept must fold. That is the vocabulary’s promise to lookup, not another display glossary.

“Same results” is testable: the sets of hit ids from the two queries should be equal. Ranking may differ slightly with match field; objects the preferred term can find must not be dropped.

Why it happens

People do not learn the preferred term before querying. They bring speech, old names, abbreviations, names in another language. If the index only stores the preferred literal, those queries all miss, and the vocabulary does not exist for searchers. Mapping lifts the query from strings to concepts, the index recalls by id, and the snippet echoes the user’s word or the preferred term. Folding must happen on the concept id: string replacement will drag in unrelated documents that happen to contain the word, and will miss objects indexed only by id, never by that literal.

Mapping without echoing makes people think they searched for A and received an unaccountable B. Writing non-preferred terms into the nav instead of mapping them breaks uniqueness on the display side.

Studying it

Pair queries and test set equality; do not stop at “the alias found a few things.”

• Paradigms: for each concept, submit the preferred term and several non-preferred terms; compare hit-id sets (precision/recall against the preferred-term baseline); also compare whether the query box was rewritten and how the snippet explains the mapping.
• Independent variables: replace versus expand versus id-fold; presence of a “searched as this concept” hint; inclusion of cross-language aliases.
• Dependent variables: Jaccard overlap with the preferred-term result, count of already-indexed objects missed, whether people understand why documents under the preferred term appeared.
• Methodological note: click-through is biased by snippet highlighting. Compare sets before ranks. A wrong mapping (two concepts’ aliases crossed) can look like healthy overlap while pointing at the wrong concept; use gold ids, not document counts.

Where it stops holding

Typos and inflection are not non-preferred terms; they belong to approximate matching or stemming and must not be smuggled into vocabulary mapping as if they were controlled. When the user is currently distinguishing two near-concepts that really differ in this system (“bill” versus “invoice”), mapping erases the distinction and the pair must be refused. Live query suggestions that silently swap an alias for the preferred term may overwrite intent that is not finished yet; mapping should run after submit, while the suggestion layer only shows the correspondence.

Applying it

• After tokenization, the retrieval pipeline looks up the non-preferred table, rewrites to a concept id, then queries the index. Keep the original string in the box; label results with the matching preferred term.
• Regression: for every core concept, compare id sets from alias queries and the preferred-term query. Overlap below threshold fails.
• Mapping table and cataloguing vocabulary share one source. Search engineering must not keep a second synonym list that will drift.
• Verify with an old name that is never written on pages. Search should list every object tagged with that concept’s preferred term—none missing, none from a neighbour concept. “Found some” is not mapping.

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 · G1.10.3 Authority files hold canonical names for entities and resolve homographs and aliases · G1.10.4 Controlled vocabularies are staff-maintained, unlike folksonomies · G1.10.6 Vocabularies need versioning and change records as the domain evolves
• Adjacent: G1.08 Terminology consistency · G3.04 Spelling correction · G3.14 Fuzzy matching and approximate search
• Search terms: non-preferred term mapping · query expansion · controlled vocabulary retrieval