Can a chain contain only one mention?

Some annotation schemes represent singleton clusters; others omit them from their task or evaluation.

Must its mentions appear in consecutive sentences?

No. References can recur after intervening material, although greater distance may make interpretation harder.

Does the chain prove the entity exists?

No. Fictional, hypothetical or mistakenly described entities can still be tracked consistently within a discourse.

AI Glossary / Natural language and speech / Intermediate

Coreference chain

Definition

A coreference chain is a set of expressions in a discourse that refer to the same entity. It can connect a name, a description and a pronoun across sentences. In NLP, the set is also called a coreference cluster; it records shared reference rather than mere similarity between words.

What it means

In “Mina opened the package. The engineer inspected it. She found a broken seal,” Mina, the engineer and She can form one chain if the context establishes that Mina is the engineer. The package and it form another. The broken seal is related to the package but is not the same object.

A chain helps a system keep track of who or what a passage discusses even when the wording changes. The word chain should not imply that each expression refers only to its immediate predecessor. All members share an entity within the chosen discourse interpretation.

The annotation scope matters. Some datasets include single mentions as one-member clusters; others score only repeated mentions. Some focus on entities, while other tasks include events. A glossary example should state its intended interpretation instead of assuming that every collection makes the same choices.

How it works

A system first identifies candidate mentions, such as noun phrases and pronouns. It then decides which mentions belong together, often using surrounding language and learned patterns. These decisions may be made jointly rather than in separate processing stages.

The resulting cluster should preserve identity consistently. If one mention refers to the same entity as a second and the second to the same entity as a third, all three belong together under an identity-based scheme. An incorrect link can therefore merge material about different entities and spread errors through a larger chain.

Shared spelling is insufficient evidence. Two people can have the same name, and two occurrences of “the manager” can refer to different people after a topic shift. Conversely, a name and a pronoun can corefer without sharing any letters. A system should distinguish discourse tracking from linking an entity to an external database identifier.

Three practical examples

A document summarizer gathers statements about one company despite alternating uses of its full name, abbreviation and “the firm.”
A support assistant tracks a particular replacement device across a customer's description without merging it with the original device.
A translation reviewer checks whether pronouns remain associated with the intended participants as names and descriptions change across sentences.

Think of it this way

Think of placing colored tabs beside all the expressions that point to one item in a story. One color follows the person, another follows the package.

The color represents shared reference, not a shared category. Two packages can look alike and still require separate colors. The analogy also assumes the links have been correctly interpreted; a mistaken tab can make unrelated statements appear to concern the same thing.

Worked example

Building two chains without merging a part and its whole

Use the invented passage: “A technician collected a tablet. The worker tested it. The device had a cracked screen.” Assume the worker refers to the technician and the device refers to the tablet.

The person chain contains two mentions: a technician and The worker. The tablet chain contains three: a tablet, it and The device. The phrase a cracked screen denotes a part of the tablet, so it is not added to the tablet's identity chain. Under a scheme retaining singletons, it forms its own one-mention cluster.

There are six listed mentions across three clusters of sizes two, three and one. The count depends on this deliberately stated mention policy. The example shows that a useful relation, such as part-of, must not be mistaken for sameness. A richer representation can record both the separate entities and their relationship.

Illustrative case study

A fictional returns team summarizes reports containing an original appliance and its replacement. An early model merges “the unit” with whichever device was mentioned most recently, even when the customer has switched back to describing the original.

Reviewers inspect chains alongside the surrounding sentences and retain unresolved cases for clarification. The team tests whether summaries assign faults to the correct device, rather than checking only whether all appliance-related sentences were collected.

The lesson is that a chain's usefulness depends on preserving entity identity. A large cluster that gathers many relevant-looking phrases can be worse than smaller correct clusters if it mixes two products. The scenario is illustrative and does not claim an observed model improvement.

Common misconceptions and mistakes

A coreference chain is a list of synonyms. Coreference depends on reference in context, not dictionary equivalence.
Identical names always belong together. Different entities can share a name.
Associated objects should be merged. A screen and its tablet are related but distinct.
Every pronoun must have a textually named antecedent. Some references depend on the situation or other context outside the passage.

Business use

Coreference chains can support summaries, document search and information extraction by organizing claims around entities. They also create a risk of transferring a statement to the wrong person, company or object when clusters are incorrectly merged.

Evaluate both mistaken merges and mistaken splits. A merge confuses different entities; a split treats repeated references to one entity as separate. Keep links inspectable through the source spans and surrounding text. If an external customer or product identifier is needed, treat that mapping as a further step with its own evidence rather than assuming that a within-document cluster establishes a real-world identity.

Check your understanding

A passage mentions “a laptop” and later “its keyboard.” Should both full phrases be placed in the same identity chain?

Answer: No. The keyboard is a part of the laptop, not the laptop itself. The possessive reference and the whole noun phrase require separate attention.

Related terms

Learn next

Mark mentions in a short paragraph and group only those that denote the same entity. Record related-but-distinct objects separately, then compare how different annotation policies treat singletons and nested mentions.

Sources

Jurafsky and Martin, Speech and Language Processing, Chapter 24: https://web.stanford.edu/~jurafsky/slp3/24.pdf. Mentions, discourse entities, coreference chains, singletons and entity linking. The technician passage and its six-mention analysis are original teaching examples with an explicit interpretation and counting policy.

Previous: Copula Next: Coreference resolution