AI Glossary / Natural language and speech / Beginner
Referent
Definition
A referent is the person, object, event, or other entity that someone refers to through an expression in context. In natural-language processing, it is often represented as an entity in a discourse model. Different expressions can have the same referent, and the same expression can have different referents on different occasions.
What it means
The expression is the wording; the referent is what that wording picks out in the intended interpretation. In “Mina entered. The technician waved,” the name and description can refer to the same person if the context identifies Mina as that technician.
A discourse model tracks such entities while a text or conversation unfolds. It can contain representations of fictional or hypothetical entities, so inclusion in that model does not prove real-world existence. Some traditions define reference directly in relation to real-world entities; this entry follows the discourse-oriented approach commonly used in computational coreference work.
How it works
When processing a passage, a system identifies relevant mentions and considers which entity each introduces or accesses. A new description may introduce an entity, while a later name or pronoun may return to one already represented. The system uses grammatical and contextual evidence to connect mentions appropriately.
Coreference resolution groups mentions that concern the same discourse entity. Entity linking takes a further step by associating a discourse entity with an external identifier or knowledge-base entry where appropriate. A successful grouping inside a document does not automatically establish which customer record or real-world person is intended.
The process also needs boundaries. Not every pronoun is a straightforward label for a particular individual. Some uses are nonreferential, and others behave like variables in quantified statements. Ambiguous expressions can leave several candidates available. An interpretable system should retain the competing candidates or request clarification when the distinction matters, rather than assuming that the nearest name is always correct.
Three practical examples
1. Document summarization: A system connects “the supplier,” “North Quay Tools,” and a later “it” when the document supports their referring to one organization.
2. Conversation support: The referent of “I” changes when the speaker changes, so a meeting assistant needs speaker context to attribute statements correctly.
3. Catalog assistance: “This one” can refer to a product selected on screen. Text alone may be insufficient if the relevant pointing or selection information is missing.
Think of it this way
Think of several labels attached to one item in a storeroom. A product name, shelf description, and inventory label can all identify that item. The analogy helps separate labels from what they identify, but language is more flexible: referents can be abstract, fictional, ambiguous, or supplied by the surrounding situation rather than by a physical tag.
Worked example
Consider this invented exchange: “Mina picked up a blue folder. She opened it.” For this exercise, we count four mentions: “Mina,” “a blue folder,” “She,” and “it.” The intended interpretation assigns Mina and She to one person, and a blue folder and it to one folder.
There are four mentions but two referents. The two coreference clusters each contain two mentions. Counting every mention as a separate entity would produce four entities, twice the intended number. Merging everything into one cluster would make the opposite mistake, confusing the person with the folder.
Now consider two separate turns: Mina says, “I found it,” and Ravi says, “I checked it.” The two occurrences of I have different referents because the speakers differ. Even identical wording therefore does not establish shared reference. These are stipulated teaching interpretations, not results from a tested coreference model.
Illustrative case study
Fictional case: Meadow Desk is testing a support assistant on equipment conversations. A customer describes a printer and a scanner, then writes, “It stopped again.” The assistant attaches the complaint to the most recently named device without checking whether the context supports that choice.
Reviewers change the workflow so that uncertain references remain unresolved until surrounding evidence or a short clarification identifies the device. The assistant can preserve both equipment records while the question is open. The case illustrates why confident language is not sufficient evidence of a correct referent. No reduction in support errors is claimed without a separate evaluation.
Common misconceptions and mistakes
Do not confuse an antecedent with a referent. An antecedent is a linguistic expression related to a later anaphoric expression; the referent is the entity those expressions concern under the intended interpretation.
Do not assume that identical names identify the same person across records, or that different descriptions must identify different people. Both decisions require contextual evidence. Similarly, a grammatical match can narrow candidates without proving identity.
Avoid forcing every expression into an ordinary entity cluster. The interpretation of quantified, nonreferential, or discourse-level expressions may need different treatment. Annotation guidelines should state which kinds of reference are included before teams compare their counts or model results.
Business use
Correct reference tracking supports reliable summaries, case histories, and conversational actions. A mistaken link can attribute a complaint to the wrong product or an instruction to the wrong participant, even when every individual word was transcribed correctly.
Businesses should test whether a system keeps distinct entities separate, connects justified repeated mentions, and exposes uncertainty when context is insufficient. For operational use, the link from a discourse entity to a database identifier deserves its own check. Understanding who a document discusses and selecting the correct account are related tasks, but success on the first does not guarantee the second.
People also ask
Can two different words have the same referent?
Yes. A name, description, and pronoun can identify the same entity in context.
Can the same word have different referents?
Yes. I changes with the speaker, and a repeated name can describe different individuals.
Must a referent be physically present?
No. Language can concern absent, abstract, fictional, or hypothetical entities, depending on the discourse and analysis.
Check your understanding
“Ravi arrived. The engineer sat down.” Must this passage contain two people?
Answer: No. Ravi and the engineer may have the same referent if context identifies Ravi as the engineer. The two expressions alone do not force either one-person or two-person interpretation.
Related terms
- Mentions
- Antecedent
- Anaphora
- Discourse model
- Entity linking
- referring expression
- mention
- coreference
- deixis
Learn next
Compare mention detection with coreference resolution and entity linking. Mark expressions in a short passage, group only those supported as sharing a referent, and leave ambiguous links explicit. Then examine how speaker changes affect first-person pronouns.
More in Natural language and speech
- Recursive transition network - A recursive transition network, or RTN, represents grammatical rules as connected state diagrams that can call one another.
- Reduced vowels - Reduced vowels are vowels pronounced with a less complete articulatory gesture than their full forms.
- Reference point - A reference point is a temporal anchor used to interpret when an event occurs relative to a viewpoint in a sentence or discourse.
- Referential density - Referential density is the degree to which participants in discourse are expressed overtly rather than left unexpressed and recovered from context.
- Referring expression generation - Referring expression generation is the task of producing language that helps a listener or reader identify an intended entity.
- Reformulation - Reformulation is expressing an earlier message again in a different form.
Sources
Source: Jurafsky and Martin, Speech and Language Processing, archived January 2025 chapter 23, introductory discourse-model discussion and section 23.1: https://web.stanford.edu/~jurafsky/slp3/old_jan25/23.pdf . Relevant primary author text was opened and read, including its distinction between discourse entities and external entity linking. Examples, counts, and the Meadow Desk case are original illustrations.
Previous: Reference text Next: Referential Constraint