Is anaphora limited to pronouns?

No. Descriptive noun phrases and, in broad NLP usage, repeated names can participate in backward reference.

Can anaphora cross sentence boundaries?

Yes. An earlier entity can remain relevant across several sentences or turns, depending on the discourse.

Is an unresolved pronoun always a system failure?

No. The available text may itself be ambiguous. A system should distinguish insufficient evidence from an avoidable processing error.

AI Glossary / Natural language and speech / Intermediate

Anaphora

Definition

Anaphora is the use of a later expression to refer back to something introduced earlier in a discourse. In “A courier arrived. She carried a parcel,” the second sentence uses anaphora to continue talking about the courier without repeating the full noun phrase. The later referring expression is an anaphor, and its earlier textual connection is an antecedent.

What it means

Anaphora lets people build connected descriptions across sentences. Readers keep track of entities and interpret expressions such as it, they, the device, or the team against that evolving context. NLP systems need similar connections when summarizing documents or answering questions.

This entry concerns discourse reference, following the supplied textbook's broad NLP terminology. In rhetoric, anaphora also names repetition at the beginnings of successive phrases or clauses. That rhetorical device is a different use of the same word and should have a separate explanation if included in the website.

How it works

A system identifies mentions and interprets whether later expressions access an existing discourse entity. It can combine syntactic information, semantic compatibility, discourse prominence, and learned contextual representations. Several mentions referring to the same entity can be grouped into a coreference chain.

The relationship is not a simple nearest-word lookup. A later pronoun may refer across intervening sentences, and multiple compatible candidates may remain. Useful processing includes the possibility that a reference is ambiguous or cannot be resolved from the available text.

Anaphora and coreference are closely related but should not be treated as identical labels for every phenomenon. Anaphora concerns a backward-looking discourse dependency; identity coreference concerns expressions referring to the same entity. Wider discourse interpretation also includes associative links and references to events or propositions, which may need different representations.

Three practical examples

1. Customer feedback: “The package arrived late. It was also wet” connects both complaints to the package, allowing a summary to preserve their shared target.

2. Meeting notes: “The finance team reviewed the proposal. They requested revisions” requires a link before the request can be attributed to the correct group.

3. Reading assistance: A tool can make a long passage easier to follow by showing which earlier entity a pronoun refers to, while marking alternatives when the text remains ambiguous.

Think of it this way

Think of discourse as a set of labeled folders that a reader opens as entities are introduced. Anaphora lets later sentences add information to an existing folder. The analogy has limits: language does not provide explicit folder identifiers, references can be uncertain, and some later expressions concern a whole event or statement rather than one object folder.

Worked example

Consider an invented text: “A technician checked the printer. The device passed the first test. It failed the second test.” The intended identity chain is {the printer, The device, It}. The two test results belong to the same printer but describe different tests, so they are not contradictory merely because one is a pass and one a failure.

Now change the first sentence to “A technician checked the printer beside the scanner.” The later expression The device could refer to either item without further context. A resolver that always selects the closest noun would choose the scanner, but proximity alone does not establish that this is the writer's intention.

Suppose an original evaluation exercise has twenty such references: twelve clearly resolved by local context and eight deliberately ambiguous. A system that correctly links the twelve and flags all eight has made twelve supported links and eight abstentions. Calling that “60 percent correct” without mentioning the ambiguity policy would be misleading: no unique gold link was specified for the remaining eight.

The evaluation must define whether it rewards identifying ambiguity, requires multiple candidates, or excludes unresolved cases. These counts are invented to illustrate evaluation design, not a reported score for an anaphora-resolution model.

Illustrative case study

Fictional case: River Documents builds summaries of equipment reports. Its assistant produces a confident fault attribution after encountering two nearby device names and an ambiguous it.

The team changes the review workflow to expose the relevant source sentences and preserve uncertainty when the target cannot be established. It tests whether the summary avoids assigning a fault to the wrong device. The case is fictional; it demonstrates why a supported unresolved result can be more faithful than a forced reference link.

Common misconceptions and mistakes

Do not assume all pronouns are anaphoric. In “It is snowing,” it does not refer back to a named object. References to something visible in the situation may also lack a textual antecedent.

Do not confuse backward reference with cataphora, where an expression precedes the later mention that clarifies it. Nor should every related pair of nouns be merged: “a car” and “the engine” can be linked through a part-whole inference without denoting the same thing. Different annotation projects may include or exclude such relationships, so their scope must be explicit.

Business use

Anaphora handling supports reliable attribution in summaries, question answering, and information extraction. The business concern is often not whether a pronoun receives a label but whether a statement is attached to the right person, product, or organization.

Evaluate on full documents and conversations that resemble the intended material. Short isolated sentences remove context that can be essential. Review mistaken merges, missed links, and unsupported certainty separately, and keep the original evidence available when a resolved reference affects a downstream action or report.

Check your understanding

“A customer entered with a colleague. They asked for a receipt.” Is they guaranteed to mean only the customer?

Answer: No. Depending on context, they may refer to both people or be a singular pronoun for one person. A confident single-entity link requires more evidence than this passage provides.

Related terms

Learn next

Compare identity coreference with bridging and cataphora. Practice annotating references in complete passages, including examples where the most responsible interpretation is a set of alternatives rather than one forced link.

Cataphora

Sources

Source: Jurafsky and Martin, Speech and Language Processing, Chapter 24, read for discourse reference, anaphora, coreference, cataphora, and related phenomena: https://web.stanford.edu/~jurafsky/slp3/24.pdf . The supplied full-book retrieval was unavailable. Examples and evaluation counts are original. This entry uses the NLP reference sense, not rhetorical repetition.

Previous: Anaphor Next: Anaphoricity detector