1 Problem definition and scope

Coreference resolution is the task of identifying which mentions in a text refer to the same underlying entity or situation. A mention is an expression that can denote something in the world of the text, such as a pronoun (“she”), a noun phrase (“the manager”), or a proper name (“Alice”). The system produces an output in which mentions are grouped into clusters representing shared referents, or—equivalently—linked through a set of coreference relations.

In practical NLP settings, coreference resolution is often treated as a structured prediction problem: the model must decide both which mentions are eligible to corefer and which prior mentions (or clusters) they should connect to. Because coreference cues arise from syntax, semantics, and discourse context, the difficulty of the task scales with document length and the complexity of the relationships between mentions.

1.1 Mentions, entities, and referents

A mention is a textual span that can participate in a coreference link. Entities are the latent objects in the meaning representation (e.g., a person, organization, or event) that mentions point to. A referent is the particular entity or situation instantiated within the document context. Coreference resolution typically aims to recover referents by clustering mentions that share them.

Different datasets adopt slightly different conventions: some define coreference in terms of a single semantic entity, while others allow links for broader eventive or nominal phenomena. Regardless of convention, successful resolution requires tracking reference over distance, handling paraphrastic descriptions, and resolving pronoun-like forms with ambiguous antecedents.

1.2 Types of coreference (e.g., nominal, pronominal, demonstrative)

Coreference phenomena vary by mention type. Nominal mentions include common nouns and their determiners (“the scientist,” “a researcher”). Pronominal mentions are pronouns and possessives (“he,” “their,” “it”). Demonstratives (“this,” “that,” “these”) often refer to nearby discourse entities or situations, and their resolution can rely strongly on discourse and attention-like cues.

Some systems distinguish also between proper-name mentions and descriptive mentions, because the matching signals differ: names may require alias handling, while descriptions require semantic compatibility and discourse context.

Coreference resolution overlaps with other reference-related tasks but differs in scope and output form. In many pipelines, these tasks are complementary: one may resolve entity identity within a knowledge base, while another groups in-text mentions that refer to the same discourse entity.

1.3.1 Coreference resolution vs. entity linking

Entity linking maps a mention to a specific entry in an external knowledge base (such as a Wikipedia page). Coreference resolution, by contrast, relates mentions to each other inside the text, without requiring the referent to be tied to a particular external identifier. In practice, entity linking may help disambiguate “John” when multiple Johns exist, whereas coreference helps determine that “he” refers to the same John previously introduced.

Because coreference can be performed without external resources, it is frequently used even when knowledge bases are unavailable. Conversely, linking can provide stable identifiers but does not inherently provide grouping structure among all mentions in the document.

1.3.2 Coreference resolution vs. anaphora resolution

Anaphora resolution is often used more broadly to describe resolving expressions that depend on prior context, including cases that are not strictly coreferential. Coreference resolution is typically a specific form of anaphora resolution where the linked expressions share the same referent. For example, some anaphoric expressions indicate a relation (like containment or part–whole) rather than identity, which may not be labeled as coreference in standard annotations.

Terminology varies across the literature, but coreference resolution is usually defined in terms of clustering mentions with shared reference, rather than resolving every context-dependent dependence type.

1.3.3 Coreference resolution vs. co-reference in knowledge graphs

Knowledge graph coreference refers to merging or aligning nodes that represent the same real-world entity across sources. While the conceptual goal—identifying whether two representations refer to the same thing—is similar, knowledge graph co-reference operates at the level of graph nodes and cross-document provenance. Textual coreference resolution is constrained by the document’s language phenomena and typically outputs within-document clusters.

As a result, knowledge graph co-reference can involve additional metadata, record linkage, and provenance modeling that are not part of the typical coreference resolution problem in NLP.

1.4 Document-level vs. sentence-level resolution

Coreference often requires information beyond a single sentence. Document-level resolution considers the full context of a text and can link mentions separated by multiple intervening clauses. Sentence-level resolution restricts the search scope to within one sentence; it is easier but misses many realistic dependencies, such as pronouns referencing earlier events or entities.

Modern systems commonly perform document-level inference, sometimes with strategies to control computational cost for long texts (e.g., pruning unlikely mentions or limiting candidate antecedents).

2 Modeling approaches

Coreference resolution systems generally share a high-level pipeline: identify candidate mentions, score possible links, and assemble the chosen links into clusters. Different modeling traditions vary mainly in how they represent candidates and how they enforce consistency across multiple decisions.

2.1 Candidate mention identification

Mention detection produces spans that could participate in coreference. Some approaches treat mention detection as a separate step trained for span identification; others use neural models that jointly infer mentions and clusters. Candidate mention identification typically includes both rule-based heuristics and learned predictors, because the number of potential spans grows quickly with document length.

Common span choices include noun phrases and pronouns; mention detection may also consider named entities and specific syntactic constructions. Errors in mention detection often propagate to clustering performance, making mention identification a key component.

2.2 Pairwise and antecedent scoring

A typical formulation scores whether a mention’s potential antecedent—an earlier mention that might be its referent—matches. For each mention, the system compares it to a set of earlier candidates and selects the best match or decides to start a new cluster.

Pairwise scoring can be implemented with neural encoders that combine representations of the candidate mention and antecedent, optionally including features for syntactic distance and semantic compatibility. A special option for “no antecedent” handles the case where the mention introduces a new referent.

2.3 Mention-ranking and span-based methods

Mention-ranking methods score and rank antecedents for each mention, then select the top scoring antecedent under certain constraints. Span-based methods, often used in end-to-end architectures, treat the mention itself as a span representation and compute scores over possible span pairs.

These methods can benefit from global context representations (e.g., from transformer-based encoders) while keeping the decision structure manageable by limiting candidate antecedents through heuristic pruning or learned filtering.

2.4 Clustering-based resolution

Instead of selecting a single antecedent per mention, clustering-based approaches build groups of mentions that share a referent. The central question becomes how to decide cluster membership and how to maintain cluster consistency.

2.4.1 Agglomerative clustering strategies

Agglomerative clustering starts from individual mentions and merges clusters iteratively based on pairwise merge criteria. The merge criterion can be a learned classifier that estimates whether two clusters should be combined. This strategy naturally produces clusters but can be sensitive to early merge mistakes.

To reduce error propagation, some systems use constraints, calibrated merge thresholds, or re-ranking steps to revise decisions. The computational cost can also increase because cluster pair comparisons may grow with the number of mentions.

2.4.2 Neural clustering objectives

Neural clustering objectives may learn to predict cluster assignments jointly, sometimes using differentiable objectives that approximate clustering decisions. A common direction is to train models with loss functions that reward correct cluster grouping and penalize incorrect merges.

These objectives can incorporate features capturing higher-order relationships implicitly (e.g., through attention over context) while still producing a discrete cluster output.

2.5 Rule-based and constraint-based methods

Rule-based systems encode linguistic heuristics such as agreement constraints (number, gender, person), syntactic patterns (e.g., pronoun antecedents in specific grammatical roles), and string matching cues (e.g., exact name matches). Constraint-based systems treat coreference decisions as satisfying a set of logical or probabilistic constraints, sometimes using optimization techniques.

While these systems may offer interpretability and can work well for narrow domains, they typically struggle with wide-ranging linguistic variation unless the rules are extensive and carefully maintained.

2.6 Hybrid systems (rules + ML/neural)

Hybrid approaches use rules to generate candidate antecedents or provide features and constraints to a statistical learner. For instance, agreement-based filters can reduce the candidate set before a neural scorer is applied. Alternatively, neural models can propose links, and a subsequent constraint solver can enforce global properties such as transitivity or cluster consistency as defined by the evaluation protocol.

Hybrid designs aim to improve robustness and efficiency by combining the precision of deterministic constraints with the generalization ability of learned models.

3 Feature representation

Feature representation defines how a model converts text into signals that predict coreference. Modern systems often rely on contextual embeddings, but many also incorporate explicit linguistic and discourse features.

3.1 Linguistic features (syntax, morphology, semantics)

Syntactic and morphological information helps determine grammatical compatibility between mentions. Dependency relations can indicate likely antecedent positions, while morphological features such as determiners and case markers can support pronoun resolution in languages where such features are explicit. Semantic features may capture type information (person, organization, location) or event relatedness.

Even when the model is neural, explicit linguistic signals can improve sample efficiency by providing structured guidance that the model might otherwise learn indirectly.

3.2 Distance and salience

Distance measures how far apart mentions are in the text, often affecting the likelihood that a later mention refers back to an earlier one. Salience reflects prominence in discourse, including factors like recency, grammatical subject status, and participation in the current focus.

Models may compute salience features explicitly (e.g., grammatical role indicators) or implicitly (e.g., via attention mechanisms that emphasize relevant context).

3.3 Agreement features (number, gender, person)

Agreement features capture morphological compatibility between pronouns and potential antecedents. English gender and number may be partially indicated by pronoun forms (“he” vs. “she,” singular vs. plural), while number agreement can often be more reliable. Some datasets also encode compatibility assumptions for reflexives and possessives.

When agreement information is missing or ambiguous (e.g., “they” used for a singular referent in some contexts), models must rely more heavily on semantic and discourse cues.

3.4 Semantic compatibility signals

Semantic compatibility assesses whether a mention and candidate antecedent are compatible in meaning. This includes type compatibility (person vs. organization), world knowledge proxies (e.g., pronoun expects an animate antecedent), and contextual role alignment (who did what in the narrative).

Neural embeddings can represent compatibility implicitly, but some systems augment them with explicit semantic-type features or lexical cues.

3.5 Contextual embeddings

Contextual embeddings encode meaning based on the surrounding text, enabling the model to distinguish between different senses and discourse roles of the same surface form. Typically derived from transformer encoders, embeddings provide rich signals for both mention representations and pairwise scoring.

3.5.1 Token-level vs. span-level representations

Token-level representations assign vectors to individual tokens, whereas span-level representations aggregate tokens within the mention span (e.g., averaging, attention pooling, or using the encoder’s start/end states). Span representations are important because coreference mentions are often multi-token phrases.

Choosing span encodings influences performance: better span pooling can improve handling of complex noun phrases and multi-word entities.

3.6 Discourse features

Discourse features capture how information is organized across sentences. They may include indicators of discourse relations, topic continuity, and cohesion patterns.

3.6.1 Discourse relations and cohesion cues

Cohesion cues include repetition of lexical items, synonyms, parallel structures, and connective phrases that signal narrative flow. Discourse relations can indicate whether an antecedent is likely to remain in focus (e.g., after an elaboration vs. a topic shift).

Some models integrate discourse structure through learned attention over discourse indicators, while others rely on the transformer’s implicit modeling of such dependencies.

4 Practical system components

Real systems must handle data preparation, inference, and output formatting with attention to edge cases. Components described here reflect common engineering patterns.

4.1 Preprocessing and normalization

Preprocessing can include sentence splitting, token normalization, and optional lemmatization. For pronoun resolution, case handling and punctuation normalization can matter because mention boundaries and tokenization affect span generation.

Normalization may also address artifacts such as extra whitespace, special character sequences, or inconsistent quote formatting present in corpora.

4.2 Tokenization and span generation

Tokenization determines how text is converted into model-ready units. Span generation enumerates candidate text spans that could be mentions, guided by syntactic parses, named entity recognizers, or span enumeration heuristics.

Because exhaustive span enumeration can be expensive, systems often restrict spans to lengths observed in training data or prioritize noun phrase boundaries and pronoun positions.

4.3 Post-processing and merging clusters

After scoring links or merge decisions, systems convert outputs into clusters. Post-processing may enforce constraints such as merging transitive links consistently or removing singleton clusters depending on the evaluation protocol.

Some pipelines also apply thresholding to decide when to link two mentions versus leaving them separate. The choice of thresholds affects the balance between precision and recall.

4.4 Handling nested or overlapping mentions

Texts can contain nested spans, such as a smaller noun phrase inside a larger one. Handling overlapping mentions requires careful decisions because assigning one span to a cluster may conflict with overlapping span assignments.

Common strategies include allowing overlaps but representing them as distinct mention nodes, using tailored mention detection to avoid certain nesting configurations, or selecting a canonical span for each coreference unit based on length and head information.

4.5 Dealing with unknown or ambiguous entities

Ambiguous entities can arise when multiple plausible antecedents exist. Systems manage this uncertainty through probabilistic scoring, confidence estimates, and learned representations that contextualize entities based on surrounding information.

When entities are unknown (e.g., rare names or out-of-vocabulary forms), contextual embeddings derived from subword modeling and document-level context can still provide useful signals, though performance may degrade compared with frequent entities.

5 Evaluation and benchmarks

Evaluation measures both the correctness of predicted coreference links and the quality of cluster structure. Because coreference is a clustering task, metrics must account for partial matches and cluster-level errors.

5.1 Metrics for coreference quality

Common metrics include precision, recall, and F1 computed over coreference links. Link-based metrics compare whether pairs of mentions that share a cluster in the gold annotation are also clustered together by the system. Cluster-aware metrics assess similarity between predicted clusters and gold clusters, often using pairwise overlap aggregations.

Some protocols emphasize mention-based link counting, while others incorporate adjustments that penalize wrong clustering more heavily than missing links.

5.2 Standard evaluation protocols

Most benchmarks follow established splits and evaluation scripts to ensure comparability. Systems are typically evaluated at the document level with the exact same preprocessing and mention boundaries as specified by the dataset guidelines.

Because datasets differ in annotation and mention definitions, models trained on one dataset may not transfer perfectly to another without adaptation.

5.3 Error analysis categories

Error analysis often categorizes mistakes into: failure to detect mentions, incorrect antecedent selection, incorrect cluster merging, and over-linking (linking mentions that should be separate). Another category is under-linking (splitting mentions that should connect).

Analyzing errors at these levels helps identify whether improvements should focus on mention detection, pairwise scoring, or clustering strategy.

5.4 Dataset characteristics and limitations

Benchmarks differ in domain (news, conversational text, fiction), annotation granularity, and linguistic coverage. Domain mismatch can lead to degraded performance when training data does not reflect the syntax and discourse patterns of the test domain.

5.4.1 Biases in training data

Training data may contain biases such as stereotyped discourse structures or limited mention types. Models can overfit to surface patterns, like pronoun position and common narrative templates, reducing their effectiveness in atypical texts.

Biases may also appear in the distribution of gendered pronouns, plural forms, or the frequency of particular syntactic constructions, influencing learned compatibility cues.

5.4.2 Annotation guidelines impact

Annotation guidelines define what counts as coreference, including whether certain bridging-like phenomena are included and how nominal anaphors are treated. Different guidelines can produce systematic differences in labels, which complicates cross-dataset comparisons.

Models trained under one labeling convention may require post-processing or fine-tuning to align with another.

6 Challenges and edge cases

Coreference resolution faces difficult linguistic situations where simple matching cues fail. Handling such cases is essential for reliable performance.

6.1 Non-referential pronouns and expletives

Some pronouns do not refer to an entity, such as expletive “it” in weather or dummy subject constructions. Since these pronouns are not coreferential with a meaningful antecedent, models must learn to recognize when linking is inappropriate.

Datasets typically annotate these cases in ways that affect metrics, so systems must follow the dataset’s treatment of non-referential uses.

6.2 Cataphora and deferred references

Cataphora occurs when a pronoun precedes its antecedent (“When she arrived, Mary…”). Many coreference models are oriented toward earlier antecedents, making cataphora challenging. Handling cataphora may require additional mechanisms that allow later antecedents during inference or special model designs.

Even when cataphora is rare, it can contribute significant errors in specific genres.

6.3 Split antecedents and collective reference

Split antecedents happen when a pronoun refers to multiple entities jointly (e.g., “They” referring to two previously mentioned people). Collective reference involves grouping that is not a single individual but a shared role or team-like unit. Resolving these requires models to represent multi-mention antecedents or cluster membership patterns beyond one-to-one linking.

Not all datasets annotate such phenomena equally, so evaluation can vary depending on labeling.

6.4 Bridging vs. coreference

Bridging involves relatedness without identity (e.g., “the door” and “the house” connected by part–whole or typical associations). Bridging may be frequent in nominal references but is often not labeled as coreference. Systems must avoid incorrectly merging entities that are merely associated.

Distinguishing identity from associative links generally requires semantic modeling and careful adherence to annotation definitions.

6.5 Reflexives and long-distance dependencies

Reflexive pronouns (“himself,” “herself”) require careful checking of binding constraints, which can span sentence boundaries in some constructions. Long-distance dependencies also appear in narratives where antecedents are far from their anaphors.

Learning binding-like constraints is especially important for languages and dataset types where reflexives are informative and frequent.

6.6 Cross-sentence ambiguity

Pronouns and descriptions can be ambiguous even with multiple candidate antecedents across several sentences. The farther the antecedent, the fewer cues remain, increasing uncertainty. Intervening events can also introduce competing entities with similar syntactic roles, leading to mislinks.

Document-level context modeling and effective candidate pruning are common strategies to mitigate cross-sentence ambiguity.

7 Applications in information processing

Coreference resolution supports downstream tasks by making discourse structure explicit. By grouping mentions that refer to the same entity, it enables consistent tracking of who or what participates in events.

7.1 Information extraction and entity consistency

In information extraction, unlinked mentions can fragment evidence about an entity. Coreference resolution helps aggregate attributes, actions, and relations tied to the same referent, improving consistency in extracted facts. It also reduces duplication by merging multiple mentions of the same entity into a single representation for the extraction module.

These benefits are particularly clear in long documents where an entity is introduced early and referenced repeatedly.

7.2 Summarization and narrative coherence

Summarization systems may generate references that depend on resolved antecedents to avoid contradictions or confusing switches in entity references. By understanding which mention corresponds to which entity, systems can choose consistent pronouns, proper names, or descriptive phrases.

Coreference-aware summarization also supports maintaining event coherence, especially in narrative text where actions are reported with different referential forms.

7.3 Question answering and retrieval augmentation

In question answering, pronouns in the question or supporting passages may need resolution to determine the intended target. Coreference resolution can rewrite passages into a more explicit form, improving retrieval and matching against query terms. It can also support multi-hop reasoning by clarifying entity trajectories across sentences.

For retrieval augmentation, resolved coreference links can be used to expand the query with aliases or related mentions, improving recall.

7.4 Dialogue systems and conversational memory

Dialogue systems require maintaining context between turns. Coreference resolution helps interpret “it,” “that,” or “they” relative to previously discussed entities, supporting more coherent conversational responses. When combined with a memory module, coreference links can update the dialogue state with consistent entity tracking.

In user-facing settings, effective coreference handling reduces misunderstandings caused by ambiguous pronouns and shifting references.

8 Resources and tooling

A range of datasets, libraries, and integration patterns exist for coreference resolution. Selection depends on language coverage, model availability, and evaluation requirements.

8.1 Common datasets and corpora

Benchmarks include newswire and other narrative corpora with expert annotations for coreference clusters. Some datasets cover multiple genres, while others focus on specific language phenomena or annotation schemes.

For researchers and practitioners, dataset choice affects both the kinds of errors likely to occur and the interpretability of results across domains.

8.2 Libraries and reference implementations

Several NLP libraries provide pretrained coreference models or utilities for running inference. Tooling commonly includes mention extraction utilities, model wrappers, and evaluation scripts consistent with benchmark protocols.

Using reference implementations can help ensure that preprocessing and output formatting match the dataset requirements, reducing evaluation discrepancies.

8.3 Integration with NLP pipelines

Coreference resolution is often integrated as a preprocessing step in larger systems. Typical integrations include converting resolved clusters into normalized text (e.g., replacing pronouns with canonical mention forms) or producing structured outputs for downstream components.

Pipeline integration must address issues like tokenization compatibility, mention boundary alignment, and efficient handling of long documents to avoid excessive latency.

9 Future directions

Ongoing research focuses on improving discourse modeling, robustness across domains and languages, and efficiency at scale.

9.1 Better discourse modeling

Future systems aim to represent discourse state more explicitly, including entities under focus, topic shifts, and coherence relations. Improved discourse modeling can help the model decide which antecedents remain salient and reduce errors from misleading superficial cues.

Hybrid approaches combining neural encoders with structured discourse representations are a common direction.

9.2 Robustness to domain shift

Models trained on one genre may degrade on another due to differences in syntax, mention style, and annotation patterns. Robustness efforts include domain-adaptive training, uncertainty-aware inference, and better calibration of link thresholds.

Evaluation under domain shift can guide improvements, emphasizing generalization rather than only in-domain performance.

9.3 Multilingual and cross-lingual coreference

Multilingual research addresses language-specific pronominal systems, agreement behavior, and mention structures. Cross-lingual transfer can leverage shared representations or parallel data to improve performance where labeled coreference data is limited.

Handling languages with rich morphology and flexible word order often requires specialized modeling choices and careful evaluation.

9.4 Efficient inference for long documents

Document-level resolution can be computationally expensive because candidate mentions and antecedent pairs grow with text length. Efficiency improvements include candidate pruning, sparse attention mechanisms, incremental processing, and more scalable clustering strategies.

Because many real-world documents are long, faster inference with minimal loss in accuracy is a major practical priority.