Is coreference resolution the same as named-entity recognition?

No. Named-entity recognition identifies and categorizes certain mentions; coreference groups mentions that share a referent.

Can it work across documents?

Yes, some tasks address cross-document coreference, but their evidence and evaluation differ from a within-document task.

Does a high-confidence prediction eliminate ambiguity?

No. Confidence is a model output and can be miscalibrated; the source passage may still support alternatives.

AI Glossary / Natural language and speech / Intermediate

Coreference resolution

Definition

Coreference resolution is the task of determining which expressions in a discourse refer to the same entity. It can link names, descriptions and pronouns into clusters, allowing an NLP system to follow people, organizations, objects or other referents across changing wording.

What it means

A user may say, “I received a replacement charger. It stopped working.” To summarize the complaint correctly, the system needs to connect It to the replacement charger. Merely recognizing charger as a product word does not supply that relationship.

The task includes more than pronouns. “Harbor Tools,” “the company” and “the supplier” may refer to one organization in a passage, while two occurrences of “the supplier” may denote different organizations in another passage. Context determines the link.

Coreference within a document also differs from identifying a record in an external database. A model can decide that several phrases concern one discourse entity without knowing that entity's customer identifier or whether every claim about it is true. External linking and factual verification require additional evidence.

How it works

A system identifies candidate mentions and decides which ones belong together. Rule-based approaches can use grammatical constraints and carefully ordered matching rules. Learned approaches use annotated examples to estimate compatibility between mentions or to build entity clusters directly.

One influential neural approach represents candidate text spans, scores whether they are mentions and scores possible antecedent links. It can also select an option representing no earlier antecedent. Pruning reduces the number of candidates that need detailed comparison. This describes a model family, not the only possible implementation.

The decisions must remain consistent at the cluster level. A wrong link can combine statements about two different entities, while a missed link fragments information about one entity. Recency and word similarity provide clues, but neither is a universal rule. Some passages genuinely lack enough information for a unique resolution, so uncertainty should remain visible.

Three practical examples

A question-answering system connects a pronoun in a retrieved paragraph with the organization named in the preceding sentence.
A complaint summarizer distinguishes an original item from a replacement, attaching each reported fault to the correct device.
A document-review tool groups references to the same project across names and descriptions while leaving uncertain links available for inspection.

Think of it this way

Think of following changing labels on folders as a discussion unfolds. Several labels may point to the same folder, but identical-looking labels can also belong to different folders.

Coreference resolution tries to establish those relationships from the discourse. The analogy does not make the system an authoritative identity checker. Its folders represent an interpretation of the text, which can be mistaken, incomplete or about fictional entities.

Worked example

Resolving what can be resolved

Consider: “A courier handed a parcel to a receptionist. The parcel was damaged. It needed repacking.” The explicit repeated description links the two parcel mentions, and the final It naturally continues that parcel reference in this constructed example.

Now replace the last two sentences with “The courier thanked the receptionist because they were helpful.” Without additional context, they can admit different readings, including a singular reference. Choosing whichever person appears nearest is not a justified universal solution.

A review workflow should distinguish the relatively clear parcel chain from the unresolved person reference. It can present candidate links or ask for clarification if a consequential action depends on the answer. The example demonstrates that a system's obligation is not to invent certainty for every pronoun; successful processing includes recognizing when the available text is insufficient.

Illustrative case study

A fictional facilities team uses AI to summarize equipment tickets. One ticket describes a damaged monitor and a replacement monitor, then says “it still flickers.” The model confidently assigns the final complaint to the replacement, but the writer's intended reference is unclear.

The team revises its review flow to display the source sentence and candidate devices when the distinction affects the next maintenance step. It also adds unambiguous examples to ensure the system is not simply flagging every pronoun as uncertain.

The lesson is to evaluate both resolution and restraint. A workflow that always guesses can misdirect work, while one that never links clear references is unhelpful. The case illustrates a design tradeoff without asserting an observed accuracy gain.

Common misconceptions and mistakes

The nearest noun is always the antecedent. Context and grammar can point elsewhere.
Matching names guarantees identity. Different people or organizations can share a name.
Occupational stereotypes are valid evidence about a pronoun's referent. They can introduce systematic errors and should not substitute for the text.
Every pronoun refers to a previously named entity. Some are nonreferential or depend on information outside the passage.

Business use

Coreference resolution is valuable when information must be assembled across several sentences. It can improve the coherence of summaries and help attribute events correctly, but errors can transfer claims between entities. The practical importance depends on what the downstream system does with those claims.

Evaluate false merges and false splits on representative documents, including multiple similar entities, corrections and long-distance references. Record whether mentions are supplied in advance or detected by the system, because these are different evaluation conditions. Inspect effects on the actual output, such as whether a summary assigns a fault or commitment correctly, rather than relying only on an aggregate benchmark score.

Check your understanding

A model links all mentions of “the manager” in a document, even though two departments each have a manager. What error should reviewers suspect?

Answer: A false merge. Shared wording is not enough to establish that the mentions denote the same person; the local and broader context must be checked.

Related terms

Learn next

Start by marking mention spans, then attempt clustering separately. Compare errors caused by missing a mention with errors caused by linking two mentions incorrectly, and include examples where the correct response is unresolved.

Sources

Jurafsky and Martin, Speech and Language Processing, Chapter 24: https://web.stanford.edu/~jurafsky/slp3/24.pdf. Task definition, mention ranking, evaluation issues and bias.

Lee, He, Lewis and Zettlemoyer, End-to-end Neural Coreference Resolution: https://aclanthology.org/D17-1018.pdf. Primary research on jointly learned mention and antecedent decisions. The examples and facilities scenario are original illustrations.

Previous: Coreference chain Next: Coreset