Can one word be a mention?

Yes. A name or referring pronoun can occupy a single word.

Can mentions overlap?

Depending on the annotation scheme, nested references can be marked; other tasks restrict the allowed spans.

Must a mention match a knowledge-base entry?

No. The referent may be new, private, fictional, or outside the database. Mention detection does not require an external identifier.

AI Glossary / Natural language and speech / Beginner

Mentions

Definition

Mentions are occurrences of expressions in a text or conversation that refer to entities, events, or other discourse referents, within a specified task's scope. Different mentions may refer to the same entity, and identical wording may refer to different entities.

What it means

A mention is the expression at a particular location, not the entity itself. In “Lena arrived. The consultant smiled,” the two expressions may refer to one person. They remain two mentions because they occur in two places.

Names, descriptions, and pronouns can all serve as mentions. An entity may be a fictional person or a hypothetical object discussed in a story or plan; being mentioned does not establish real-world existence. Equally, a company name appearing in a sentence does not establish that the company performed every action described nearby.

Annotation projects define their coverage differently. A system designed for named company references may omit ordinary objects, while a broader coreference task includes them. Event mentions and references to whole statements require additional choices beyond a simple inventory of people and things.

How it works

A text-processing system records each mention's position, often with start and end offsets, alongside its words and optional type. Detection proposes the spans. Coreference resolution groups mentions that share a referent. Entity linking may connect a mention to an external knowledge-base record.

These are separate operations even when one model learns them jointly. A mention can be detected without being resolved, and a resolved document entity may have no matching external record.

Boundaries and nesting need explicit rules. In “Ravi's bicycle,” a scheme may mark the whole bicycle phrase and the embedded person reference. A mention that appears only once is a singleton. Excluding singletons from a particular benchmark is an annotation choice, not evidence that those expressions have no meaning.

Three practical examples

Customer service: “my tablet,” “the device,” and “it” may be three mentions of one item in a complaint.
News search: two occurrences of “Jordan” need context because they may identify a person, a country, or different people.
Project notes: “the launch” and “it” may refer to an event, which should not automatically enter a database restricted to physical assets.

Think of it this way

Think of mentions as labels placed at different points in a narrative, with threads connecting labels that identify the same subject. Labels are the occurrences; the connected subject is the entity. The analogy is limited because a thread can remain uncertain, and some expressions refer to an entire situation rather than a neatly named object.

Worked example

Use the passage “Lena ordered a printer. The consultant installed it. Lena tested the device.” Assume the context states that Lena is the consultant and that only one printer is under discussion. Under a policy marking full referring noun phrases and pronouns, there are six mentions.

The person cluster contains “Lena” in sentence one, “The consultant,” and “Lena” in sentence three. The printer cluster contains “a printer,” “it,” and “the device.” Thus there are six occurrences, five distinct surface strings, and two entities.

The repeated name counts twice because it occurs at two different positions. Removing the contextual assumption about the consultant could make the person clustering uncertain without changing the six detected expressions. Counting mentions and deciding identity are different tasks, so a dashboard should not label six as the number of people or objects.

Illustrative case study

Fictional case: Cedar Search builds a report index that displays how often a supplier is discussed. Its first dashboard counts unique spelling forms, treating a full name, an abbreviation, and “the supplier” as separate organizations.

An editor requests separate measures for mention occurrences and resolved organizations. The revised design stores the original span, the tentative document entity, and any verified business identifier in distinct fields. Ambiguous descriptions remain unresolved. This is an illustrative data-design correction, not a claim that automatic linking has become error-free.

Common misconceptions and mistakes

Do not treat every noun phrase as a straightforward referring mention under every scheme. Expressions can be generic, predicative, negated, or otherwise handled differently by annotation policies. The “it” in “It is snowing” is not a reference to a previously discussed machine.

Do not confuse a mention with an endorsement, verified fact, or unique person. Repeating a name does not establish that all occurrences denote the same individual. Conversely, a change from a name to a pronoun does not necessarily introduce someone new.

Business use

For business analysis, decide what the count represents before designing charts. Mention frequency measures textual attention under the detection policy; it does not directly measure customers, incidents, sentiment, or market share.

Keep source positions and document identifiers so reviewers can inspect each occurrence. If text is edited after offsets are created, maintain a version or regenerate the offsets. Review false merges and missed links separately, especially when summaries associate responsibility, ownership, or complaints with a named organization. Aggregate only after the relevant identity assumptions have been checked.

Check your understanding

A report uses “Delta Works” four times, and every occurrence refers to the same company. How many mentions and entities are represented by those occurrences?

Answer: Four mentions and one entity, assuming the identity interpretation is correct. A count of one distinct spelling form answers a third, different question.

Related terms

Learn next

highlight the mentions in a short business message and count occurrences before grouping them. Mark unresolved identity separately. Then compare document-level coreference with entity linking, which asks whether a reference corresponds to a specific external record.

Entity linking

Sources

Sources: Jurafsky and Martin, Speech and Language Processing, Chapter 24, mentions and coreference scope: https://web.stanford.edu/~jurafsky/slp3/24.pdf . Soon, Ng, and Lim (2001), coreference markables and identity of reference: https://aclanthology.org/J01-4004.pdf . Lee et al. (2017), text spans and antecedent assignments: https://aclanthology.org/D17-1018.pdf . The counts and business scenario are original examples.

Previous: Mention-pair Next: Mercer's theorem