Does the detector identify the antecedent?

Not necessarily. A binary detector only decides whether the mention is anaphoric; a separate or joint component selects an antecedent.

Is a first mention nonreferential?

No. It can refer to an entity while being discourse-new and therefore lacking an earlier textual antecedent.

Must anaphoricity detection be a separate stage?

No. A model can learn it jointly with mention detection and coreference decisions.

AI Glossary / Natural language and speech / Intermediate

Anaphoricity detector

Definition

An anaphoricity detector estimates whether a mention in a text refers back to an entity already introduced in that discourse. It asks whether an earlier antecedent is needed or available, rather than directly deciding which earlier mention is correct. The task can be implemented as a separate classifier or integrated into a larger coreference-resolution model.

What it means

In “A printer arrived. The machine is damaged,” the machine is anaphoric under the intended same-printer reading. By contrast, the first mention a printer introduces an entity rather than referring back. A later pronoun does not automatically qualify: it in “It is raining” is nonreferential in that construction.

Anaphoricity is distinct from referentiality. A first mention can refer to a real or imagined entity while having no earlier antecedent. A detector that labels all non-anaphoric mentions “meaningless” would therefore confuse two different questions. The exact labels also depend on the discourse boundaries and annotation policy.

How it works

A separate detector takes a candidate mention and its context, then returns a label or score. Training examples come from annotated text that marks which mentions refer back. Features or learned representations can capture wording, grammatical form, position, and surrounding discourse.

One possible use is filtering candidates before antecedent selection. However, a hard filter can create unrecoverable mistakes: if it removes a truly anaphoric mention, the later resolver cannot link it. Joint models can instead consider mention status and antecedent alternatives together, including an option representing no appropriate earlier antecedent.

A non-anaphoric mention may still need to remain in the discourse model as a potential antecedent for future mentions. Excluding it from a backward-link decision should not mean deleting the entity from all later processing. This distinction is especially important when a pipeline separates detection, linking, and entity tracking.

Three practical examples

1. Support conversations: A detector can help distinguish a newly introduced replacement device from a later reference to the original device, reducing inappropriate automatic merging.

2. Document extraction: A pipeline can identify which noun phrases require a backward reference decision and which introduce new entities that should remain available for future links.

3. Pronoun analysis: A model can distinguish referential it from nonreferential constructions before attempting to attach every occurrence to a nearby noun.

Think of it this way

Think of a receptionist deciding whether an arrival belongs to an existing file or needs a new file. That preliminary decision is separate from selecting the exact existing file. The analogy has limits: language can be ambiguous, reference can depend on unstated context, and a model's score is not the same as verified identity. New files must also remain available for later arrivals.

Worked example

Consider an invented evaluation set of 100 candidate mentions. Annotators mark 40 as anaphoric under a stated identity-reference policy. At one threshold, a detector predicts 30 positive mentions, of which 24 are truly anaphoric.

There are 24 true positives, 6 false positives, 16 false negatives, and 54 true negatives. Precision is 24/30 = 80 percent; recall is 24/40 = 60 percent. If every negative prediction is permanently excluded from linking, sixteen real backward references have already been lost before antecedent selection begins.

At a lower threshold, suppose 50 mentions are predicted positive, with 36 true positives. Precision is 36/50 = 72 percent, and recall is 36/40 = 90 percent. This retains more real anaphors but passes fourteen false positives to later processing.

Neither threshold is declared best by these numbers alone. The downstream resolver's behavior and the cost of missed links versus incorrect merges matter. A correct positive label also does not guarantee that the right antecedent will be selected. These counts are a constructed teaching example, not benchmark results or evidence for a particular classifier.

Illustrative case study

Fictional case: Elm Records adds a separate anaphoricity filter to an entity-extraction pipeline. It reduces the number of candidate links, but reviewers find that some first mentions have disappeared entirely, leaving later references with nothing to link to.

The team separates “do not seek an earlier antecedent for this mention” from “discard this mention.” It evaluates both the detector and final entity chains, and compares hard filtering with a joint decision approach. The fictional case describes a modeling issue without claiming a real performance gain.

Common misconceptions and mistakes

Do not assume that definite noun phrases always refer back. A phrase beginning with the can introduce an entity identifiable from shared knowledge or situation. Similarly, not every pronoun is anaphoric.

Do not confuse anaphoricity accuracy with coreference accuracy. Correctly deciding that a reference exists leaves the antecedent-selection problem unsolved. Also avoid treating all unannotated spans as reliable negatives without checking the dataset: some corpora omit singletons or particular reference types, so absence of a label can reflect annotation scope rather than a linguistic fact.

Business use

Anaphoricity detection can make the internal decisions of a document-understanding pipeline easier to inspect. It helps teams identify whether errors arise from missing a reference, inventing one, or selecting the wrong antecedent.

Evaluation should use the intended document boundaries and representative genres. A mention can be new within one extracted paragraph but already introduced in the full document. Report precision and recall together with downstream linking outcomes, and retain ambiguous cases rather than forcing a misleading binary gold label where the evidence does not support one.

Check your understanding

A detector correctly flags it as anaphoric but links it to the wrong one of two devices. Has the full reference task succeeded?

Answer: No. The anaphoricity decision was correct, but antecedent selection failed. The pipeline should evaluate and diagnose those decisions separately as well as together.

Related terms

Anaphora
Mention detection
Coreference resolution
Precision
Recall
referentiality
discourse-new mention
antecedent selection

Learn next

Compare a hard-filter pipeline with a joint mention-ranking model that includes a no-antecedent option. Trace what happens to first mentions, singleton entities, ambiguous references, and false-negative anaphors.

False negative

Sources

Source: Jurafsky and Martin, Speech and Language Processing, Chapter 24, Section 24.3 and architecture discussion, read for anaphoricity classifiers, features, filtering trade-offs, and joint modelling: https://web.stanford.edu/~jurafsky/slp3/24.pdf . The supplied full-book retrieval was unavailable. Confusion-matrix values and the fictional case are original. Dataset-specific annotation scope must be checked before defining training negatives.

Previous: Anaphora Next: Ancestral graph