Can an entry contain several words?

Yes. A gazetteer can include names and expressions spanning multiple words, subject to the matcher being used.

Is a gazetteer a machine-learning model?

Not necessarily. It can be a manually maintained list applied by a deterministic matcher, though its results can feed a learned model.

Does a matched name identify one unique entity?

Not always. Shared names and aliases can require additional entity-disambiguation or linking steps.

AI Glossary / Natural language and speech / Beginner

Gazetteer

Definition

A gazetteer in natural-language processing is a collection of names or terms used to help recognize matching expressions in text. Entries may be grouped by type, such as places, organizations, products, or dates. The word can also refer to the lookup component that applies those lists. A match provides useful evidence, but does not by itself resolve the expression's meaning in context.

What it means

The traditional geographical sense concerns place names. NLP uses the term more broadly for typed lists that support text analysis. A product-name gazetteer, for instance, might contain official names and selected aliases, with associated identifiers or categories.

Recognizing a string and identifying a real-world entity are separate steps. “Jordan” can be a person's name or a country, while an ordinary word may also be a brand. List membership indicates a possible interpretation; the surrounding sentence may be needed to decide which interpretation applies.

How it works

A team prepares entries and their associated information, then configures how the lookup operates. Matching choices can include case sensitivity, word boundaries, normalization, and handling of overlapping entries. Implementations differ, so these details should be specified rather than assumed.

When the matcher finds an entry, it can annotate the corresponding text span and attach its list category or other metadata. Later rules or a learned model may use that annotation alongside contextual evidence. Some systems use gazetteer membership as one feature rather than accepting every match as a final entity label.

Maintenance is part of the method. New names, renamed products, aliases, spelling variants, and outdated entries affect coverage and precision. A larger list can find more genuine mentions while also adding ambiguous strings that create false positives. Evaluation must therefore consider both what is found and what is incorrectly labeled.

Three practical examples

1. Product-support search: A maintained list of product names and approved aliases helps locate mentions in incoming messages, with context used to resolve ambiguous names.

2. Place-name extraction: A system marks potential city names in documents, then uses surrounding information to distinguish locations sharing the same name.

3. Document annotation: A research team uses a gazetteer to suggest candidate organization spans for review, reducing repetitive lookup work while retaining human correction.

Think of it this way

Think of a guest list checked against names heard at an event. A matching name tells you that someone with that name is on the list, but it does not establish that the speaker means that exact person.

The analogy is limited: NLP entries can be multiword phrases, aliases, or non-person entities, and matching can involve normalization. A gazetteer is also not necessarily a complete or authoritative registry of everything that exists.

Worked example

Build a fictional product gazetteer containing “Nova” and “Nova Pro,” both marked as candidate product names. For this exercise, use case-sensitive whole-word matching and retain overlapping matches for later review.

In “Nova Pro replaces Nova,” the matcher finds three spans: Nova Pro at the beginning, Nova within that longer name, and the final Nova. If a documented longest-match policy suppresses contained matches, two spans remain: the full Nova Pro and the final Nova.

Now consider “Nova submitted the order,” where context establishes Nova as a person's name. The gazetteer still finds the string Nova, but treating the list category as unquestionable would produce a false product label. Contextual disambiguation is necessary.

Finally, “nova pro” produces no match under the exercise's case-sensitive policy. Switching to case-insensitive matching could recover it, but would also change which other strings are considered. The example yields different results because of explicit policy choices, not because the abstract word gazetteer determines one universal matching behavior.

Illustrative case study

Fictional case: An office-supply company maintains a list of product families for support-message routing. A newly added product shares its name with a common first name. The router begins sending some customer-account questions to the product team.

Reviewers inspect matched spans and context, then introduce a disambiguation step and tests for both product and personal-name uses. They also document how overlapping names are handled. The company keeps the list's useful coverage while recognizing that a vocabulary update can change false-positive behavior.

Common misconceptions and mistakes

A gazetteer is not restricted to geographic names in NLP. Its entries can represent many entity categories, although the intended scope should be clear.

Absence from a list does not prove an expression is not an entity. The list may be incomplete, outdated, or missing an alias. Presence likewise does not prove the category is correct for that occurrence.

Exact matching is not the only possible implementation, and whole-word or longest-match behavior should not be assumed. A system's configuration determines how variants and overlaps are treated. A gazetteer also does not automatically perform entity linking to a unique knowledge-base record.

Business use

For businesses, gazetteers can make domain vocabulary explicit and reviewable. Assign an owner for updates, keep the source of entries, and test additions against realistic text rather than checking only that the new name can be found.

Measure missed mentions and incorrect matches separately. Preserve the original span and the reason for a lookup suggestion so reviewers can inspect errors. When names are ambiguous, combine list evidence with context and a defined uncertainty policy before routing requests or populating final records.

Check your understanding

Why does the fictional longest-match policy leave two product-name spans in “Nova Pro replaces Nova” rather than three?

Answer: It suppresses the shorter Nova span contained within Nova Pro, while retaining the separate final Nova. That is an explicit overlap policy, not a universal property of all gazetteers.

Related terms

Named entity recognition
Entity linking
Lexicon
alias
string matching
text annotation
precision and recall

Learn next

Learn next about named entity recognition and entity disambiguation. Build a tiny list with an alias, an overlapping name, and an ambiguous ordinary word, then document how the matching policy affects each case.

Named entity recognition

Sources

GATE's official gazetteer chapter was read for typed lists, lookup annotations, multiword matches, configuration differences, and downstream contextual rules: https://gate.ac.uk/sale/tao/splitch13.html . The Nova examples deliberately specify their own matching policies and are not presented as default GATE behaviour. The product-routing case is fictional, and no real customer data or product list was processed.

Previous: Gaussian scale mixture Next: Gelbach decomposition