Is this the same as naming an entity?

Naming is one option. The task also includes descriptions and, in broader generation settings, pronouns chosen for the context.

Must a description mention every known property?

No. It should provide useful identifying information. Listing everything can burden readers and distract from the relevant distinction.

Can an expression be unique but unhelpful?

Yes. A hidden serial number might distinguish one object in a database while giving a person no usable way to locate it.

AI Glossary / Natural language and speech / Intermediate

Referring expression generation

Definition

Referring expression generation is the task of producing language that helps a listener or reader identify an intended entity. A system might choose “the blue folder,” “Mina,” or “it,” depending on the entity, available context, and audience. The challenge is to say enough for successful identification without making the wording unnecessarily difficult.

What it means

The system has a target and decides how to describe it for someone else. Reference resolution works in the opposite direction: receiving an expression and determining its referent.

Success is contextual. “The folder” may work when only one folder is relevant, while “the blue folder beside the printer” may be needed among several. A property stored in a database is useful only if the audience can recognize or access it. A hidden inventory number may uniquely identify an object internally but fail as an instruction to a visitor.

How it works

A generation process can start with a target entity, a set of potential distractors, and information about properties and relationships. Content selection chooses identifying information, and linguistic realization turns that information into an expression. Some systems also decide among names, descriptions, and pronouns using discourse context.

Classical algorithms often build a distinguishing description: one that is true of the target and excludes relevant alternatives. An incremental method can examine properties in a preferred order, adding useful ones until the target is distinguished. It need not find the mathematically shortest description.

Relationships can also identify a target, such as its position relative to another object. The other object must itself be understandable, and careless recursive descriptions can become circular. Learned methods can use visual and language representations, but they still need evaluation against the intended referent and surrounding context. Fluent wording alone is insufficient.

Three practical examples

1. Equipment guidance: An assistant says “the upper blue connector” when “the connector” would leave several choices, provided the user can see the relevant color and position.

2. Document discussion: A meeting assistant introduces “the revised budget spreadsheet” and later uses a shorter reference when the conversation makes the intended file clear.

3. Image interaction: A system describes a selected object within an image so that another person can identify it among nearby objects.

Think of it this way

Think of writing a label for a parcel on a shared shelf. The label must distinguish the parcel from its neighbors using information the recipient understands. The analogy captures identification and audience needs. Natural referring expressions also depend on conversation history, grammar, and perception, which a fixed parcel label may not capture.

Worked example

Consider an original scene with four objects: A is a small blue folder, B a large blue folder, C a small red folder, and D a small blue notebook. The target is A. Assume these properties are accurate and visible to the reader.

“The folder” leaves A, B, and C. Adding blue leaves A and B. Adding small leaves only A. Thus “the small blue folder” distinguishes the target in this context.

Check each selected property. Without small, B remains. Without blue, C remains. Without folder, D remains. All three are needed to distinguish A using these simple properties in this scene.

Now remove B from the scene. “The blue folder” already distinguishes A; small is no longer required for uniqueness. It could still help a reader notice the target, but that is a separate usability question. The example demonstrates contextual discrimination, not a universal rule that longer or shorter expressions are always better.

If the display hides color, the system cannot simply reuse the description and assume success. It needs another accessible property, a clear relation, or an interactive way to indicate the target.

Illustrative case study

Fictional case: A repair tutorial tells users to “remove the fastener.” During testing, several users hesitate because two fasteners are visible. The development team changes the instruction to identify the intended fastener using a stable position relative to a clearly labeled component.

Reviewers then test the wording from the user's viewpoint, including a rotated diagram. They discover that an unqualified “left” can change with perspective, so the final instruction specifies the relevant view. The case illustrates why database uniqueness and human identification must be checked together; it does not report a measured safety improvement.

Common misconceptions and mistakes

A common mistake is assuming the shortest expression is automatically best. A little redundancy may make identification easier, while an extremely compact technical label may confuse the intended audience.

Another is choosing a true property that does not distinguish the target. If every relevant folder is blue, blue does not reduce the competitor set, even though it remains accurate.

Pronouns also require care. “It” may be grammatical yet ambiguous when two plausible referents are active. Finally, descriptions should not depend on invisible information, unstable spatial relations, or circular references such as identifying one unknown object only through another equally unknown object.

Business use

Businesses encounter this task in generated reports, digital assistance, training instructions, and interfaces that describe visual content. The practical goal is successful identification by the intended audience in the actual setting.

Test whether readers select the correct item, how long they take, and where they request clarification. Include scenes with similar distractors and changes of viewpoint. Where source data or perception is uncertain, avoid generating a precise-looking property as if it were confirmed. A wrong descriptive detail can actively direct attention to the wrong object.

Check your understanding

In the original four-object scene, why is “the small blue thing” insufficient?

Answer: Both A, the folder, and D, the notebook, are small and blue. The category folder excludes D and is needed for this description to distinguish A.

Related terms

Referent
Natural language generation
Referring expression
reference resolution
discourse context
entity description
visual grounding

Learn next

Create a small set of objects with overlapping properties. Choose a target and remove distractors one property at a time. Then change what the listener can see and decide whether the same description still works. Compare content selection with the separate task of producing grammatical wording.

Sources

Reiter and Dale, A Fast Algorithm for the Generation of Referring Expressions: https://aclanthology.org/C92-1038.pdf

Dale and Haddock, Generating Referring Expressions Involving Relations: https://aclanthology.org/E91-1028.pdf

Yu and colleagues, Modeling Context in Referring Expressions, research abstract: https://arxiv.org/abs/1608.00272

Primary passages read; examples are original.

Previous: Referential transparency Next: Reflection