Hubbry Logo
CoreferenceCoreferenceMain
Open search
Coreference
Community hub
Coreference
logo
7 pages, 0 posts
0 subscribers
Be the first to start a discussion here.
Be the first to start a discussion here.
Contribute something
Coreference
Coreference
Coreference
View on Wikipedia
from Wikipedia

In linguistics, coreference, sometimes written co-reference, occurs when two or more expressions refer to the same person or thing; they have the same referent. For example, in Bill said Alice would arrive soon, and she did, the words Alice and she refer to the same person.[1]

Co-reference is often non-trivial to determine. For example, in Bill said he would come, the word he may or may not refer to Bill. Determining which expressions are coreferences is an important part of analyzing or understanding the meaning, and often requires information from the context, real-world knowledge, such as tendencies of some names to be associated with particular species ("Rover"), kinds of artifacts ("Titanic"), grammatical genders, or other properties.

Linguists commonly use indices to notate coreference, as in Billi said hei would come. Such expressions are said to be coindexed, indicating that they should be interpreted as coreferential.

When expressions are coreferential, the first to occur is often a full or descriptive form (for example, an entire personal name, perhaps with a title and role), while later occurrences use shorter forms (for example, just a given name, surname, or pronoun). The earlier occurrence is known as the antecedent and the other is called a proform, anaphor, or reference. However, pronouns can sometimes refer forward, as in "When she arrived home, Alice went to sleep." In such cases, the coreference is called cataphoric rather than anaphoric.

Coreference is important for binding phenomena in the field of syntax. The theory of binding explores the syntactic relationship that exists between coreferential expressions in sentences and texts.

Types

[edit]

When exploring coreference, numerous distinctions can be made, e.g. anaphora, cataphora, split antecedents, coreferring noun phrases, etc.[2] Several of these more specific phenomena are illustrated here:

Anaphora
a. The musici was so loud that iti couldn't be enjoyed. –The anaphor it follows the expression to which it refers (its antecedent).
b. Our neighborsi dislike the music. If theyi are angry, the cops will show up soon. – The anaphor they follows the expression to which it refers (its antecedent).
Cataphora
a. If theyi are angry about the music, the neighborsi will call the cops. – The cataphor they precedes the expression to which it refers (its postcedent).
b. Despite heri difficulty, Wilmai came to understand the point. – The cataphor her precedes the expression to which it refers (its postcedent)
Split antecedents
a. Caroli told Bobi to attend the party. Theyi arrived together. – The anaphor they has a split antecedent, referring to both Carol and Bob.
b. When Caroli helps Bobi and Bobi helps Caroli, theyi can accomplish any task. – The anaphor they has a split antecedent, referring to both Carol and Bob.
Coreferring noun phrases
a. The project leaderi is refusing to help. The jerki thinks only of himselfi. – Coreferring noun phrases, whereby the second noun phrase is a predication over the first.
b. Some of our colleagues1 are going to be supportive. These kinds of people1 will earn our gratitude. – Coreferring noun phrases, whereby the second noun phrase is a predication over the first.

Relation to bound variables

[edit]

Semanticists and logicians sometimes draw a distinction between coreference and what is known as a bound variable.[3] Bound variables occur when the antecedent to the proform is an indefinite quantified expression, e.g.[4][clarification needed]

  1. Every studenti has received hisi grade. – The pronoun his is an example of a bound variable
  2. No studenti was upset with hisi grade. – The pronoun his is an example of a bound variable

Quantified expressions such as every student and no student are not considered referential. These expressions are grammatically singular but do not pick out single referents in the discourse or real world. Thus, the antecedents to his in these examples are not properly referential, and neither is his. Instead, it is considered a variable that is bound by its antecedent. Its reference varies based upon which of the students in the discourse world is thought of. The existence of bound variables is perhaps more apparent with the following example:

  1. Only Jacki likes hisi grade. – The pronoun his can be a bound variable.

This sentence is ambiguous. It can mean that Jack likes his grade but everyone else dislikes Jack's grade; or that no one likes their own grade except Jack. In the first meaning, his is coreferential; in the second, it is a bound variable because its reference varies over the set of all students.

Coindex notation is commonly used for both cases. That is, when two or more expressions are coindexed, it does not signal whether one is dealing with coreference or a bound variable (or as in the last example, whether it depends on interpretation).

Coreference resolution

[edit]

In computational linguistics, coreference resolution is a well-studied problem in discourse. To derive the correct interpretation of a text, or even to estimate the relative importance of various mentioned subjects, pronouns and other referring expressions must be connected to the right individuals. Algorithms intended to resolve coreferences commonly look first for the nearest preceding individual that is compatible with the referring expression. For example, she might attach to a preceding expression such as the woman or Anne, but not as probably to Bill. Pronouns such as himself have much stricter constraints. As with many linguistic tasks, there is a tradeoff between precision and recall. Cluster-quality metrics commonly used to evaluate coreference resolution algorithms include the Rand index, the adjusted Rand index, and different mutual information-based methods.

A particular problem for coreference resolution in English is the pronoun it, which has many uses. It can refer much like he and she, except that it generally refers to inanimate objects (the rules are actually more complex: animals may be any of it, he, or she; ships are traditionally she; hurricanes are usually it despite having gendered names). It can also refer to abstractions rather than beings, e.g. He was paid minimum wage, but didn't seem to mind it. Finally, it also has pleonastic uses, which do not refer to anything specific:

  1. It's raining.
  2. It's really a shame.
  3. It takes a lot of work to succeed.
  4. Sometimes it's the loudest who have the most influence.

Pleonastic uses are not considered referential, and so are not part of coreference.[5]

Approaches to coreference resolution can broadly be separated into mention-pair, mention-ranking or entity-based algorithms. Mention-pair algorithms involve binary decisions if a pair of two given mentions belong to the same entity. Entity-wide constraints like gender are not considered, which leads to error propagation. For example, the pronouns he or she can both have a high probability of coreference with the teacher, but cannot be coreferent with each other. Mention-ranking algorithms expand on this idea but instead stipulate that one mention can only be coreferent with one (previous) mention. As a result, each previous mention must be given a score and the highest scoring mention (or no mention) is linked. Finally, in entity-based methods mentions are linked based on information of the whole coreference chain instead of individual mentions. The representation of a variable-width chain is more complex and computationally expensive than mention-based methods, which lead to these algorithms being mostly based on neural network architectures.

See also

[edit]

Notes

[edit]

References

[edit]
Revisions and contributorsEdit on WikipediaRead on Wikipedia

Coreference

View on Grokipedia
from Grokipedia
Coreference is a linguistic in which two or more expressions within a text or , such as phrases, pronouns, or proper names, refer to the same real-world , , object, or event, thereby establishing referential identity across mentions. This relation is essential for cohesion, allowing speakers or writers to avoid by reusing references to previously introduced entities, and it operates on a continuum from full identity to partial overlaps influenced by context and . In linguistic theory, coreference encompasses various subtypes, including anaphora, where a subsequent expression (the anaphor) links back to an earlier antecedent, and cataphora, a forward-pointing reference resolved by later context. Additional forms involve discourse deixis, which references prior segments of the text, and predicative relations, where expressions attribute properties to the same entity. Theoretical models, such as those drawing on mental space theory, explain shifts in coreference through cognitive operations like specification (adding details), refocusing (changing perspective), and neutralization (reducing specificity), highlighting its dynamic and context-dependent nature. Within (NLP), coreference resolution denotes the computational task of automatically detecting and clustering these coreferential mentions to enhance text understanding. Originating in the with early heuristic systems, the field advanced through annotated corpora from initiatives like the Message Understanding Conferences (MUC) and Automatic Content Extraction () in the , enabling supervised approaches such as mention-pair models and ranking algorithms. Key challenges persist, including ambiguity in non-referential phrases, handling singletons (entities with only one mention, comprising 60-86% of cases), and domain-specific issues like or temporal variations in specialized texts. Applications span , , summarization, and clinical , where accurate resolution improves system performance despite ongoing needs for robust evaluation metrics like B³ for .

Fundamentals

Definition

Coreference is a linguistic relation in which two or more noun phrases (NPs) or other expressions within a refer to the same real-world entity, abstract concept, or . This relation enables efficient communication by linking expressions that share referential identity, such as a proper name and a subsequent . For instance, in the sentence "John entered the room. He sat down," the "He" corefers with the proper name "John," both denoting the same individual. Central to coreference are referring expressions, which include proper names (e.g., "John"), definite NPs (e.g., "the man"), and pronouns (e.g., "he"), that point to entities in the discourse model. These expressions are distinguished by their roles: an antecedent is the initial expression that introduces or establishes the entity, while subsequent coreferents refer back to it. Key terminology includes "mention," which denotes any surface-level expression (typically an NP) that refers to an entity; "entity," the underlying object, person, or concept being referenced; and "coreferent chain," a sequence of two or more mentions linked as referring to the identical entity. Coreference plays a vital in maintaining coherence, allowing speakers and writers to avoid by substituting full NPs with shorter forms like pronouns, thereby facilitating smoother text flow and reader comprehension. This mechanism supports the construction of a shared of the discourse, where entities persist across sentences. Coreference encompasses subtypes such as anaphora (backward reference) and cataphora (forward reference), though these are explored in greater detail elsewhere.

Historical Development

In the early , , pioneered by , shifted focus to synchronic analysis of language as a system of signs, laying foundational ideas on how signifiers relate to signified entities and influencing later explorations of referential relations in . The mid- marked a pivotal advancement through , as integrated coreference into syntactic theory during the 1950s and 1960s. In works like (1957) and Aspects of the Theory of Syntax (1965), linked coreference to and transformational operations, introducing referential indices to track co-referential elements across sentences and distinguishing them from bound variables. This approach treated coreference as a derivational phenomenon, assuming syntactic identity often aligned with semantic equivalence, though challenges from quantifiers soon highlighted limitations. From the 1970s to the 1980s, attention turned to and , with scholars like Edward L. Keenan examining definite descriptions and their role in establishing coreference. Keenan's 1971 work examined two kinds of in , including those carried by definite descriptions regarding existence and uniqueness. This period solidified coreference's place in semantic theory, as seen in Peter Sells' 1985 Lectures on Contemporary Syntactic Theories, which synthesized anaphora across frameworks like government-binding theory and generalized . By the 1990s, coreference transitioned into with the rise of (NLP), shifting from to applied systems for resolving references in text. Early supervised methods, such as those using on annotated corpora, emerged around 1995, enabling automated clustering of mentions and marking a key evolution toward practical discourse understanding.

Types and Examples

Anaphoric Coreference

Anaphoric coreference, commonly referred to as anaphora, occurs when a linguistic expression, such as a or definite , follows an antecedent in the and derives its interpretation from that prior element. In this forward-referring relation, the anaphor points back to the antecedent to establish identity of , enabling efficient communication by avoiding repetition of full descriptions. For instance, in the sentence "The chased the . It was fast," the "it" serves as the anaphor referring to "the " as its antecedent, though ambiguity may arise if "it" could plausibly refer to "the " in context, highlighting the need for pragmatic resolution to disambiguate coreferent from non-coreferent interpretations. Linguistic constraints on anaphora are primarily governed by binding theory, a framework in that regulates the structural relations between antecedents and anaphors. Principle A of binding theory stipulates that anaphors, such as reflexives (e.g., "himself") or reciprocals (e.g., "each other"), must be bound to a c-commanding antecedent within a local domain, typically the same , ensuring the antecedent structurally dominates the anaphor. For example, in "John saw himself," "himself" is bound to "John" because "John" c-commands it from a higher structural position, whereas "*Himself saw John" violates Principle A due to the lack of such binding. These principles prevent illicit coreference, such as in cases where a pronoun cannot bind to a non-c-commanding element, distinguishing grammatical anaphora from ungrammatical attempts at coreference. In , anaphora plays a crucial role in maintaining cohesion by linking sentences and reducing redundancy, allowing speakers to track entities across a without restating them explicitly. This mechanism fosters textual unity, as seen in extended narratives where repeated pronouns or definite descriptions signal continuity of reference, enhancing readability and flow. Anaphora thus contributes to the overall coherence of by presupposing shared of antecedents, which listeners or readers recover to interpret subsequent expressions. Anaphora varies between surface and deep forms, distinguished by whether the anaphoric process relies on syntactic structure or deeper semantic interpretation. Surface anaphora, such as simple pronominal reference, requires a linguistically overt antecedent and is controlled by syntactic rules, whereas deep anaphora involves pragmatic or interpretive reconstruction without strict syntactic parallelism. A classic example of deep anaphora is (VP) ellipsis, as in "John likes apples, and Mary does too," where the elided VP under "does" is interpreted as coreferent with "likes apples" through semantic recovery rather than surface form. This distinction underscores how anaphora can operate at different levels of linguistic processing to achieve coreference. In contrast to cataphora, which involves backward reference to a forthcoming antecedent, anaphora is inherently forward-directed in its textual progression.

Cataphoric Coreference

Cataphoric coreference, or cataphora, occurs when a linguistic expression, such as a , precedes and refers forward to a subsequent antecedent that provides its full interpretation. This contrasts with the more prevalent anaphoric coreference, where reference points backward to a prior expression. Cataphora is less common than anaphora due to the cognitive processing demands it imposes on language users, as resolving the reference requires holding the initial expression in until the antecedent appears. Syntactically, it is constrained and typically occurs in specific structural contexts, such as preposed subordinate clauses or lists, where the forward reference can be anticipated within a bounded scope. For instance, in relative clauses introduced early, cataphora facilitates integration without violating locality principles. A classic example is the sentence "When he arrived, John was tired," where the pronoun "he" cataphorically refers to the later noun "John." In more complex sentences, such as those involving embedded clauses, cataphora can lead to ambiguous readings if multiple potential antecedents follow, as in "If she wins the award, Mary will celebrate with her team," where "she" anticipates "Mary" but could momentarily suggest another entity. In , cataphora serves anticipatory functions by building or structuring , particularly in stylistic writing like or technical descriptions, where it signals upcoming details to engage the reader. For example, novelists use it to character introductions, enhancing cohesion. Identifying cataphora presents challenges due to its higher relative to anaphora, as the lack of an immediate antecedent increases the risk of misresolution during comprehension. Corpus analyses of literary texts, such as those in the Anaphoric Treebank, reveal that cataphoric instances are rarer and often context-dependent, complicating and analysis in extended narratives like novels.

Theoretical Aspects

Relation to Bound Variables

In formal semantics, coreference refers to the relation where two expressions denote the same in the , establishing referential identity independent of syntactic structure. In contrast, bound variables involve scope-dependent assignments where a or anaphor is interpreted as a variable governed by a quantifier or lambda operator, as seen in frameworks like and predicate logic. This distinction is crucial because bound variables do not imply coreference; instead, they allow the pronoun's interpretation to covary with the quantifier's scope without referring to a fixed individual. A key example illustrates this difference: in the sentence "John loves his wife," the pronoun "his" is coreferential with "John," denoting the same specific individual. However, in "Every man loves his wife," "his" is not coreferential with "every man" but functions as a bound variable, interpreted such that each man in the domain loves his own wife, with the pronoun's reference varying across instances. This bound reading arises from the quantifier's scope, avoiding a unified referential link. This theoretical framework originates in , developed in the 1970s, which integrates syntax with to handle quantification and pronouns. Montague's approach treats pronouns systematically as bound variables when under quantifier scope, using translations that embed them within lambda abstractions or predicate logic operators, without invoking coreference. For instance, the of "Every man loves his " can be represented as: ∀x [man(x)→∃y [wife(y,x)∧love(x,y)]]\forall x \, [\text{man}(x) \to \exists y \, [\text{wife}(y,x) \land \text{love}(x,y)]] Here, the pronoun "his" corresponds to the bound variable xx, scoped by the universal quantifier, demonstrating how binding operates without referential identity to a single . This representation highlights the scope dependency, where the existential quantifier for the is subordinated to the universal over men. The implications of this separation are significant for semantic interpretation, as conflating coreference with binding can lead to errors in resolving scope ambiguities or quantifier interactions. By distinguishing referential links from operator scope, ensures compositional meanings that accurately capture variable interpretations in quantified contexts, influencing broader formal semantic theories.

Coreference in Formal Semantics

In formal semantics, coreference is modeled through dynamic frameworks that represent how utterances update a shared discourse context, tracking entities and their relations across sentences. Discourse Representation Theory (DRT), introduced by Hans Kamp in 1981, provides a foundational approach by constructing Discourse Representation Structures (DRS) to handle anaphoric relations. In DRT, coreferring expressions link to the same discourse referent, a variable representing an entity in the mental model of the discourse. This allows for systematic resolution of pronouns and definite descriptions by embedding conditions that equate or subordinate referents. DRT employs a notation to visually represent these structures, where referents are listed above a horizontal line and conditions below it. For instance, the "John entered the . He sat down" begins with a DRS for the first sentence introducing referent xx with the condition John(x)∧entered(x,room)John(x) \land entered(x, room): xJohn(x)∧entered(x,room)\begin{array}{c} x \\ \hline John(x) \land entered(x, room) \end{array} The second sentence updates this DRS by adding referent yy with conditions he(y)he(y) and y=xy = x, along with sat(y)sat(y), yielding: x,yJohn(x)∧entered(x,room)∧he(y)∧y=x∧sat(y)\begin{array}{c} x, y \\ \hline John(x) \land entered(x, room) \land he(y) \land y = x \land sat(y) \end{array} This equivalence y=xy = x captures coreference, ensuring the pronoun "he" refers back to John. A related framework, File Change Semantics developed by Irene Heim in 1982, treats the discourse context as a "file" of indexed entries for entities, where coreferents update the same file card rather than introducing new ones. In this system, anaphors succeed only if their descriptive content matches an existing entry, promoting incremental context updates. Advanced mechanisms in these theories address unresolved coreference via accommodation, as elaborated in Heim's 1983 work on projection. Accommodation permits the addition of presupposed referents to the when they are not explicitly introduced, enabling felicitous coreference in underspecified discourses—for example, inferring a shared acquaintance in "His is waiting" without prior mention. This process ensures semantic coherence by globally or locally adjusting the file to satisfy referential presuppositions. Cross-linguistically, formal semantic models like DRT adapt to variations in systems, such as in null-subject languages like Spanish, where pro-drop allows omitted subjects to corefer via implicit without overt forms. In Spanish, null subjects preferentially resume topic-continuous antecedents, modeled by restricting in the DRS to high-salience positions, unlike in non-pro-drop languages requiring explicit for the same links. This variation highlights how morphological options influence coreference resolution within unified dynamic frameworks.

Computational Approaches

Coreference Resolution Task

Coreference resolution is a fundamental task in (NLP) that involves identifying all mentions—linguistic expressions such as noun phrases—in a document that refer to the same real-world entity and partitioning them into coreference chains. For example, in the sentence "John entered the room. He sat down," the chain would group "John" and "he" as referring to the same individual. The goal is to produce an output that accurately clusters these mentions while handling variations in form, such as pronouns, definite descriptions, or proper names. The task typically breaks into subtasks, beginning with mention detection, which identifies candidate spans in the text that could serve as referring expressions. This step is crucial as a prerequisite, since coreference linking operates on detected mentions, and errors here propagate to subsequent clustering. Following detection, the core linking subtask groups compatible mentions into equivalence classes, effectively resolving which referents are identical without linking to external knowledge bases in the standard setup. Key datasets have standardized evaluation and training for the task. OntoNotes, introduced in the mid-2000s, provides a large-scale, multilingual corpus with coreference annotations integrated alongside other layers like and , covering genres such as news and conversational text. The CoNLL-2011 shared task established benchmarks using OntoNotes data, focusing on unrestricted coreference across English, Arabic, and Chinese, and promoting consistent annotation schemes that include both nominal and pronominal mentions. These resources shifted the field toward end-to-end systems that jointly handle mention detection and resolution on diverse, real-world texts. Evaluation metrics assess the quality of predicted coreference chains against gold-standard annotations, emphasizing in linking mentions. The Message Understanding Conference (MUC) metric, introduced in 1995, measures link-based performance by counting the number of correctly merged or split coreference sets, rewarding systems that avoid spurious merges while penalizing missed connections. B-Cubed, proposed in 1998, is a mention-centric approach that computes for each individual mention based on the proportion of correctly clustered co-referents in its entity, then averages across all mentions for an overall F1 score. The Constrained Entity-Alignment F-measure (CEAF), from 2005, aligns predicted and gold chains via a bipartite matching to optimize entity overlap, providing a more robust measure that accounts for both mention boundaries and partition structure. Shared tasks like CoNLL-2011 often report the average of these three F1 scores as a composite metric to balance their perspectives. Challenges in the task setup arise from inherent linguistic ambiguities, particularly in long-distance coreference where mentions are separated by intervening text, making contextual dependencies harder to capture. Nested entities, such as a embedded within a larger (e.g., "the man's mother" where "man" and "mother" may corefer differently), complicate mention detection and partitioning by introducing overlapping spans that defy simple linear clustering. These issues underscore the need for models robust to structural complexity in annotation guidelines like those in OntoNotes.

Algorithms and Models

Early methods for coreference resolution relied on rule-based systems that leveraged syntactic parsing to identify antecedents, particularly for . A seminal example is Hobbs' algorithm, introduced in , which operates on surface parse trees by traversing the syntactic structure from the pronoun upward and then downward to find potential antecedents based on recency and grammatical constraints. This approach, developed in the and refined through the , achieved notable success in pronoun resolution without requiring deep semantic analysis, serving as a baseline for decades due to its simplicity and efficiency. In the 2000s, statistical approaches shifted the paradigm toward data-driven models, using classifiers to predict coreference links between noun phrases. A foundational work is Soon et al. (2001), who employed maximum entropy classifiers trained on annotated corpora like MUC-7, incorporating features such as agreement in number, gender, and proper names, along with distance metrics; this system achieved approximately 60.4% F1 on the MUC-7 benchmark. These methods improved over rule-based systems by learning from examples, often outperforming them on diverse texts, though they were limited by pairwise decisions that required post-processing for full chains. The advent of neural networks in the late introduced end-to-end models that jointly predict mentions and coreference clusters without explicit pairwise . Lee et al. (2017) proposed the first such system, using a encoder to score potential antecedent spans for each mention, trained to maximize the marginal likelihood of gold clusters; on the OntoNotes benchmark, it attained 67.2% average F1, surpassing prior statistical methods by eliminating reliance on hand-crafted features or parsers. Building on this, BERT-based models from 2018 onward integrated architectures for contextual embeddings, enabling more nuanced span prediction. For instance, Joshi et al. (2019) fine-tuned BERT as a coreference head atop span representations, yielding 76.9% F1 on OntoNotes and highlighting BERT's ability to capture long-range dependencies. Advanced techniques have further refined these neural foundations, incorporating graph-based partitioning to cluster mentions into chains and specialized embeddings for better span handling. Denis and Baldridge (2010) introduced partitioning for end-to-end resolution, where mentions form vertices and edges encode compatibility scores, allowing of clusters in a single step. Integration of contextual embeddings like (in extensions of Lee et al., 2018, reaching 70.4% F1 on OntoNotes) and SpanBERT ( et al., 2020, achieving 79.6% F1) has enhanced representation of variable-length spans, with SpanBERT's pretraining on masked spans proving particularly effective for coreference by emphasizing contiguous text units. Performance trends reflect this evolution, with systems advancing from around 60% F1 in the 2000s on benchmarks like MUC to over 80% in the on OntoNotes, driven by neural architectures and pretrained models that better model context; recent models like Maverick (Xia et al., 2024) achieve 83.6% F1 using efficient DeBERTa-based pipelines, with ongoing advances incorporating large models as of 2025.

Applications and Challenges

Practical Applications

Coreference resolution plays a pivotal role in by linking multiple mentions of the same , thereby enhancing the accuracy of entity recognition and relation extraction across diverse texts. In search engines, it improves query understanding by resolving ambiguous references, such as pronouns in user inputs, allowing systems to better interpret intent and retrieve relevant results. In , coreference resolution ensures coherence by correctly linking pronouns and entities across languages, particularly in neural MT systems where or structure differences can lead to errors. By identifying and preserving coreferential relationships in source texts, it enables translators to maintain entity consistency in outputs, such as resolving "it" to the appropriate antecedent in translations from English to gendered languages like French or German. This integration has been shown to boost translation quality in document-level models. Dialogue systems benefit from coreference resolution through improved tracking of referents in multi-turn conversations, enabling chatbots to respond contextually to user mentions. Coreference helps resolve ellipses and anaphora in spoken dialogues, such as linking "that one" to a previously mentioned , which enhances natural interaction and reduces misunderstanding in voice assistants. Coreference chains support text summarization by consolidating scattered mentions of entities into unified representations, allowing for more concise and coherent abstracts while preserving key identities. This is particularly useful in abstractive summarization, where unresolved references could lead to fragmented or repetitive outputs. In domain-specific applications, coreference resolution aids legal document analysis by linking entities and events in contracts, facilitating and compliance checks through datasets like LegalCore. Similarly, in biomedical , it connects mentions of proteins, genes, and diseases across scientific texts, improving extraction from literature such as abstracts in the GENIA corpus.

Open Challenges

Despite significant progress in coreference resolution, theoretical gaps persist, particularly in handling implicit coreference that requires inferring connections from world knowledge beyond explicit textual cues. These implicit relationships challenge formal models, as they demand integration of external , which current semantic frameworks struggle to incorporate systematically. Similarly, introduces unresolved issues in aligning coreferential elements across text and , where fine-grained cross-modal associations and inherent ambiguities in visual descriptions hinder accurate resolution. For instance, resolving a in text to an depicted in an accompanying often fails due to limited annotated multimodal datasets and the complexity of encoding inter-modal dependencies. Computationally, coreference resolution faces challenges from imbalanced training , leading to lower performance on less common types or patterns. Long-distance references pose another limitation, as models degrade in accuracy when antecedents and anaphors span large textual distances, complicating dependency modeling over extended contexts. In low-resource languages, performance drops further owing to insufficient labeled and linguistic disparities that generic models cannot adapt to effectively. Cross-lingual coreference faces substantial hurdles from variations in linguistic structures, notably in pro-drop languages like Spanish or , where omitted s necessitate implicit inference not captured by annotation schemes designed for non-pro-drop languages. This is exacerbated by the scarcity of diverse, multilingual datasets that adequately represent such patterns, limiting the development of robust approaches. Ethical concerns arise from biases in coreference models, which can amplify societal stereotypes, particularly in resolution where training data imbalances lead to preferential linking of occupational roles to male entities. For example, models may erroneously corefer "" with "he" more often than "she," perpetuating disparities in downstream applications like text summarization. Looking ahead, integrating large language models (LLMs) such as the GPT series offers promising directions for zero-shot coreference resolution, enabling inference without task-specific fine-tuning, though post-2020 advances reveal persistent issues like and suboptimal handling of complex chains. These developments underscore the need for hybrid approaches that combine LLM capabilities with specialized modules to address lingering gaps in robustness and inclusivity. Historical undercoverage of cataphora, where references precede antecedents, exemplifies a broader theoretical oversight in .

References

  1. https://stel2.ub.edu/cba2010/phd/phd.pdf
  2. https://web.stanford.edu/class/archive/cs/cs224n/cs224n.1162/handouts/cs224n-lecture10-coreference.pdf
  3. https://www.sciencedirect.com/science/article/pii/S153204641100133X
  4. https://web.stanford.edu/~jurafsky/slp3/23.pdf
  5. https://aclanthology.org/J01-4004.pdf
  6. https://pmc.ncbi.nlm.nih.gov/articles/PMC2121325/
  7. https://aclanthology.org/C69-7001.pdf
  8. https://people.umass.edu/partee/RGGU_Web_12/materials/RGGU127.pdf
  9. https://www.researchgate.net/publication/220874023_Supervised_Noun_Phrase_Coreference_Research_The_First_Fifteen_Years
  10. https://plato.stanford.edu/entries/anaphora/
  11. https://people.umass.edu/partee/MGU_Web_11/materials/MGU1110_2up.pdf
  12. https://scholar.harvard.edu/files/ctjhuang/files/1983_binding_note.pdf
  13. https://udel.edu/~bruening/Downloads/ReflSUNYHO.pdf
  14. https://www.media.mit.edu/gnl/discourse/summaries/byhh.html
  15. https://www.uni-saarland.de/fileadmin/upload/lehrstuhl/steiner/Projekte/GeCo/GECo_AP2.pdf
  16. https://babel.ucsc.edu/~hank/Hankamer&Sag76.pdf
  17. https://home.uchicago.edu/merchant/pubs/surface.deep.anaphora.hd.2up.pdf
  18. https://linguistics.berkeley.edu/~mikkelsen/papers/ac_postproceedings.pdf
  19. https://www.crissp.be/wp-content/uploads/2023/01/HEY_final_version.pdf
  20. https://glossary.sil.org/term/cataphora
  21. https://www.researchgate.net/publication/371118136_On_the_Motivations_and_Pragmatic_Functions_of_Cataphora_in_Natural_Conversation
  22. https://www.researchgate.net/publication/232833689_Processing_Differences_for_Anaphoric_and_Cataphoric_Pronouns_Implications_for_Theories_of_Discourse_Processing
  23. https://www.sciencedirect.com/science/article/abs/pii/S0010027707002715
  24. https://languagelog.ldc.upenn.edu/nll/?p=2891
  25. https://www.thoughtco.com/what-is-cataphora-grammar-1689829
  26. https://www.researchgate.net/publication/10931382_The_Influence_of_Morphological_Information_on_Cataphoric_Pronoun_Assignment
  27. https://ccsenet.org/journal/index.php/ells/article/download/0/0/48816/52606
  28. https://www.sfu.ca/~mtaboada/docs/publications/Trnavac_Taboada_cataphora.pdf
  29. https://core.ac.uk/download/pdf/196591426.pdf
  30. https://semantics.uchicago.edu/kennedy/classes/s09/experimentalsemantics/grodzinsky-reinhart93.pdf
  31. https://plato.stanford.edu/entries/montague-semantics/
  32. https://www.cs.rhul.ac.uk/~zhaohui/montague73.pdf
  33. https://plato.stanford.edu/entries/discourse-representation-theory/
  34. https://oxfordre.com/linguistics/display/10.1093/acrefore/9780199384655.001.0001/acrefore-9780199384655-e-636
  35. https://aclanthology.org/H05-1004/
  36. https://aclanthology.org/N06-2015/
  37. https://cogcomp.seas.upenn.edu/papers/MentionDetection.pdf
  38. https://aclanthology.org/W11-1901/
  39. https://aclanthology.org/M95-1005/
  40. https://arxiv.org/pdf/2211.04428
  41. https://www.utdallas.edu/~sanda/courses/Discourse/Papers/Hobbs_resolving_pronouns.pdf
  42. https://aclanthology.org/J01-4004/
  43. https://aclanthology.org/D17-1018/
  44. https://aclanthology.org/D19-1588/
  45. https://aclanthology.org/C10-1017/
  46. https://arxiv.org/abs/1802.05365
  47. http://nlpprogress.com/english/coreference_resolution.html
  48. https://arxiv.org/pdf/2407.21489
  49. https://aclanthology.org/2023.eacl-main.69/
  50. https://aclanthology.org/2022.starsem-1.30/
  51. https://aclanthology.org/W99-0211.pdf
  52. https://aclanthology.org/2025.findings-acl.1284/
  53. https://aclanthology.org/2021.crac-1.2/
  54. https://w.sentic.net/survey-on-coreference-resolution.pdf
  55. https://cdn.aaai.org/ocs/10418/10418-46098-1-PB.pdf
  56. https://aclanthology.org/2023.emnlp-main.682/
  57. https://www.mdpi.com/1099-4300/26/6/529
  58. https://www.researchgate.net/publication/338993761_Coreference_Resolution_Toward_End-to-End_and_Cross-Lingual_Systems
  59. https://aclanthology.org/2025.crac-1.9.pdf
  60. https://direct.mit.edu/coli/article/47/3/615/102824/Toward-Gender-Inclusive-Coreference-Resolution-An
  61. https://arxiv.org/abs/1804.06876
  62. https://hal.science/hal-04747642v1/document
Add your contribution
Related Hubs
Contribute something
User Avatar
No comments yet.