1 Overview and definitions

1.1 What counts as anaphora

Anaphora is a language phenomenon in which an expression is interpreted by reference to something else—typically an earlier item in the same text (an antecedent) or an entity introduced through discourse context. In ordinary communication, speakers and writers often use pronouns, noun phrases, or even larger referring expressions to avoid repeating the same form. Anaphora therefore serves a functional role: it ties together successive parts of a discourse so that listeners or readers can track “who” or “what” is being discussed.

The key diagnostic is interpretation dependence. The anaphoric expression does not contribute its full referential identity on its own; instead, its intended meaning is recovered through a relationship with a previously established referent or discourse representation.

1.2 Anaphora vs. other reference types

Anaphora is frequently contrasted with related forms of reference, including:

  • Deixis, where interpretation depends mainly on situational context (e.g., “this,” “that” tied to speaker location or perception).
  • Cataphora, where a later expression supplies information needed for an earlier anaphoric-looking item (the direction of dependency is reversed).
  • Exophoric reference, where the intended referent is outside the discourse and recovered from the world, rather than from a prior textual antecedent.

These distinctions are not always absolute in practice, but they help determine whether the primary dependency is intra-textual (anaphora) or context-world based (deixis/exophora).

1.3 Antecedents and reference relations

An anaphoric expression is usually linked to an antecedent, the discourse element that supplies the interpretive content. The reference relation can be characterized in several ways, depending on analysis goals. For some theories, the antecedent and anaphor are treated as co-referential (aiming at the same entity). For other perspectives, the relation may involve looser connections, such as expressing a part–whole relationship, a set–member relation, or an association that makes the anaphor interpretable even without strict identity.

In discourse analysis, antecedents are often treated as elements of a dynamic discourse model: later expressions “see” earlier ones through memory-like activation governed by salience, coherence, and grammatical constraints.

2 Types of anaphora

2.1 Pronominal anaphora

Pronominal anaphora occurs when the anaphoric expression is realized by pronouns or pronominal determiners.

2.1.1 Personal pronouns

Personal pronouns (“he,” “she,” “they,” and analogous forms) commonly depend on a previously mentioned noun phrase or an established discourse participant. Interpretation typically draws on features such as person, number, and gender (where relevant), as well as discourse salience. For example, when a narrative introduces two characters, later pronouns require disambiguation by resolving which character remains the most salient candidate.

2.1.2 Demonstratives and determiners

Demonstratives (“this,” “that,” “these,” “those”) and determiners used anaphorically may point to an entity by using prior discourse information. Depending on the language and construction, they can behave like anaphoric devices when their interpretation is anchored in previously established referents rather than in immediate perception.

2.2 Nominal anaphora

Nominal anaphora involves anaphoric expressions that are noun phrases rather than pronouns.

2.2.1 Repetition and synonymy

A frequent strategy is repeating a noun phrase or using a closely related phrase to refer back to the same discourse entity. Synonymy and paraphrastic descriptions can also support backward reference, though analysts distinguish between cases where the expressions refer to the same individual versus cases where the relation is merely associative. Repetition tends to preserve explicit links, while synonymy relies more heavily on semantic compatibility and discourse coherence.

2.2.2 Definite descriptions

Definite descriptions (e.g., “the man,” “the decision,” “the earlier plan”) are often interpreted anaphorically when the description uniquely identifies the antecedent in the discourse context. In many frameworks, definite noun phrases are not always strictly anaphoric, but when their content is anchored in a prior introduction, they function as a common realization of nominal anaphora.

2.3 Verbal and clausal anaphora

Beyond noun phrases, anaphoric dependencies can involve verbs, verb phrases, or whole clauses.

Ellipsis is often analyzed as a form of anaphoric interpretation, especially VP ellipsis, where a verbal phrase is omitted but interpreted through an antecedent phrase. The “backing” antecedent provides the structure needed to reconstruct the missing material. Although ellipsis is sometimes treated as distinct from canonical anaphora, it shares the core property of interpretive dependence on a prior discourse element.

2.3.2 Sentence-level reference

Some constructions allow a later clause to refer to the informational content of an earlier clause or discourse segment. This can occur via pro-form-like elements (such as “so” in certain uses) or via clause-sized determiners and connectors that retrieve a proposition as an interpretation target.

2.4 Zero anaphora (implicit reference)

Zero anaphora refers to an argument position that is not overtly realized, yet is interpreted as linked to a discourse antecedent.

2.4.1 Context-driven identification

In languages where omission is licensed by grammar or discourse norms, listeners infer the missing participant or relation by relying on context. Even when no overt pronoun appears, the interpretation depends on antecedent material or established discourse roles.

2.4.2 Discourse-role recovery

Resolution often relies on discourse role assignment: who is the agent, patient, topic, or otherwise prominent participant in the relevant stretch of discourse. As a result, zero anaphora tends to be constrained by both syntactic licensing and discourse structure.

3 Anaphora in discourse analysis

3.1 Cohesion and discourse structure

Anaphora is central to cohesion, the network of links that makes separate clauses form a coherent unit. By tying later material to earlier mentions, anaphoric expressions reduce redundancy and provide a mechanism for maintaining continuity. In discourse structure models, anaphora can also serve as evidence for boundaries of segments such as episodes, topic spans, or coherence relations.

3.2 Tracking reference across turns

In spoken interaction, anaphoric dependencies often cross turn boundaries. Speakers interpret pronouns and nominal references by using what has been recently said, what has been taken up as current topic, and what participants are actively attending to. Turn-taking can introduce delays: the antecedent may be earlier in the conversation, requiring memory management and attention-based salience.

3.3 Topic continuity and shift

Anaphora interacts with topic continuity. When the discourse maintains the same topic, anaphoric reference is often easier: the antecedent remains active. When the discourse shifts topics, anaphors may either signal a return to an earlier topic or indicate a new referent that is inferable from the shifted context. Analysts therefore examine whether the anaphor points to an antecedent within the current topic domain or across a larger transition.

3.4 The role of shared context

Interpretation also depends on shared context, including background knowledge and discourse conventions. Even when an antecedent is not explicitly repeated, shared assumptions can make certain referents accessible. This role is especially visible in cases where anaphors rely on associations rather than direct identity links.

4 Mechanisms for resolving anaphora

4.1 Surface cues and grammatical agreement

A major class of resolution mechanisms uses surface cues, including agreement features (number, gender, person), syntactic position, and grammatical roles. These cues narrow the candidate antecedents. For example, mismatched features typically rule out certain antecedents, while structurally prominent phrases (such as subjects or topicalized items) often become preferred targets.

4.2 Semantic compatibility

Even when grammatical constraints allow multiple candidates, resolution often requires semantic compatibility. The anaphor’s lexical requirements and selectional tendencies can rule out antecedents that are incompatible with the anaphor’s meaning. Semantic compatibility also interacts with world knowledge and conventional associations.

4.3 Structural constraints and salience

Resolution is sensitive to salience, the degree to which an antecedent is prominent in the discourse model. Structural constraints—such as proximity, grammatical prominence, and hierarchical relationships—help determine which earlier mention remains active. Many accounts treat salience as a combination of accessibility and coherence-related factors.

4.4 Disambiguation strategies in interpretation

When multiple candidates remain plausible, interpreters use additional strategies:

  • Proximity-based preferences, favoring nearer antecedents.
  • Coherence-driven selection, aligning the anaphor with the antecedent that best supports the discourse interpretation.
  • Reanalysis and repair, where the listener revises assumptions when new information arrives.

These strategies allow incremental interpretation, where resolution can evolve as the discourse continues.

5 Theoretical approaches

5.1 Rule-based accounts

Rule-based approaches treat anaphora resolution as the application of explicit constraints and decision rules. Such systems specify which grammatical and semantic factors must hold and which candidates are ranked higher based on formal criteria. While rule-based models can be transparent, they often need extensive stipulation to capture the full range of discourse effects.

5.2 Centering theory and discourse salience

Centering theory models discourse as maintaining a set of potential discourse centers and tracking how the “current center” changes across sentences. Under this view, pronouns and other referring expressions are linked to the discourse’s accessibility structure: forms are chosen to reflect how smoothly the discourse can continue from one centered interpretation to the next.

5.3 Dynamic and model-theoretic perspectives

Dynamic approaches treat reference as incrementally updated in a discourse model. Each utterance modifies an information state, and anaphoric expressions are resolved by consulting that evolving state. Model-theoretic perspectives similarly emphasize how meanings are computed relative to representations that accumulate over time, enabling a principled treatment of variable introduction and later retrieval.

5.4 Computational and formal methods

Computational treatments build systems that predict antecedent choices using features derived from syntax, semantics, and discourse structure. Formal methods often define constraints on possible interpretations and use logic-like representations of discourse dependencies. Together, these approaches support large-scale analysis and evaluation using annotated corpora.

6 Resolution challenges and common failure modes

6.1 Ambiguity and multiple plausible antecedents

Ambiguity arises when more than one discourse entity is compatible with the anaphor’s grammatical features and semantic needs. Pronouns are especially vulnerable when several candidates are equally salient (e.g., coordinated noun phrases or multiple competing subjects). In such cases, resolution depends on fine-grained discourse cues and often remains uncertain.

6.2 Cataphora vs. anaphora confusions

Because both anaphora and cataphora involve interpretive dependence within discourse, automated or manual annotation may confuse the direction of dependency. Determining whether an earlier expression is resolved by a later antecedent (cataphora) or by something earlier (anaphora) requires careful attention to the temporal order of cues that license interpretation.

6.3 Bridging vs. direct anaphora

Not all backward reference is identity-based. Bridging anaphora involves an anaphor whose antecedent is related through a part–whole link, set membership, causal or functional association, or another inferable relationship. These cases can be challenging because the antecedent may not match the anaphor’s descriptive content directly, and resolution often depends on pragmatic inference rather than strict sameness.

6.4 Repair, rephrasing, and discourse correction

In real discourse, reference can be corrected. Speakers may rephrase after producing a problematic pronoun or omit a clarifying description later. Such repairs complicate analysis because earlier expressions may be retrospectively reassessed: what was initially taken as a resolution target may no longer be the correct antecedent once the discourse is updated.

7 Illustrative examples and analysis workflow

An analysis workflow typically begins by marking the anaphoric expression and identifying candidate antecedents. Annotators often use a consistent set of guidelines defining what counts as an acceptable antecedent relation. The goal is to capture the intended interpretive dependency rather than merely surface repetition.

7.2 Identifying boundaries of the referring domain

Next, analysts determine the referring domain, the span of discourse within which antecedents are considered eligible. This step helps prevent incorrect links to unrelated earlier mentions. Referring domains are influenced by discourse structure, topic spans, and syntactic embedding, depending on the chosen annotation scheme.

7.3 Evaluating candidate resolutions

Candidates are evaluated using grammatical constraints, semantic plausibility, and discourse coherence. When multiple links could be defended, the workflow may rank candidates by salience or select the interpretation that best supports a coherent reading of the surrounding context.

7.4 Reliability and inter-annotator considerations

Because anaphora resolution involves interpretive judgments, datasets must address reliability. Inter-annotator agreement is affected by ambiguity, guideline clarity, and the granularity of decisions (e.g., whether bridging is included, or how strictly co-reference is required). Training and adjudication protocols are commonly used to maintain consistency.

8 Cross-linguistic perspectives (non-exhaustive)

8.1 Language-specific pronominal systems

Languages vary in how they encode person, number, gender, honorifics, or definiteness, which influences anaphora. Some languages have rich agreement morphology that provides strong cues for resolution, while others rely more on discourse structure and word order. Pronoun inventories also differ in how they distinguish referents and maintain role-based tracking.

8.2 Constraints in pro-drop languages

In pro-drop languages, grammatical systems may permit omission of subjects or other arguments. This can make zero anaphora more common and shift the resolution burden from pronoun form to syntactic licensing and discourse-role interpretation. Identifying antecedents therefore involves reconstructing hidden arguments based on surrounding clauses.

8.3 Differences in definite description use

Definite descriptions may behave differently across languages in terms of when they are licensed and how they interact with discourse accessibility. Some systems use definite descriptions as a stable anaphoric mechanism, while others prefer alternative strategies such as demonstratives or morphological marking. These differences affect how analysts define and annotate nominal anaphora.

8.4 Discourse conventions affecting interpretation

Conversational norms and stylistic practices can shape anaphoric choices. For instance, some communities may favor explicit naming over pronominal reference, while others allow more frequent reliance on implicit links. Such conventions influence both the distribution of anaphoric forms and the strategies listeners use to resolve them.

9.1 Co-reference and coreference chains

Co-reference refers to the relation where expressions are interpreted as referring to the same entity. In annotation tasks, anaphoric expressions often contribute to building coreference chains, though some anaphors (especially bridging or associative ones) may not always meet strict identity criteria depending on the scheme.

9.2 Deixis and contextual reference

Deixis involves reference anchored to the communicative situation rather than to a prior textual antecedent. While anaphora is centered on intra-discourse dependencies, deixis uses features such as speaker viewpoint, spatial location, or discourse time, which can overlap with anaphoric-like uses in certain constructions.

9.3 Ellipsis and implicit anaphora

Ellipsis omits material that is recovered from context, and this recovery can be anaphoric in nature when it depends on an earlier linguistic antecedent. More broadly, implicit anaphora covers cases where reference is not fully realized by overt expressions but is inferred through discourse structure and shared expectations.

9.4 Anaphoricity in rhetorical structure

Anaphoric relations also appear in connections between rhetorical units. A later segment may refer back to a prior claim or textual span, supporting the rhetorical organization of the discourse (e.g., explanation, contrast, or elaboration). In this way, anaphora can serve not only referential goals but also contribute to the coherence of larger rhetorical structures.