Cross-Document Coreference Resolution

• Cross-document coreference resolution is defined as the process of identifying and clustering mentions across documents that refer to the same underlying entity, event, or concept.
• It encompasses varied formulations including event-centric, hierarchical, and cross-domain settings, each with unique annotation schemes and clustering strategies.
• Key challenges such as lexical variation, heterogeneous datasets, and evaluation realism drive ongoing research in model design and dataset unification.

Searching arXiv for recent and foundational CDCR papers to ground the article. Cross-document coreference resolution (CDCR) is the task of identifying and linking mentions that refer to the same underlying entity, event, concept, or scientific concept across multiple documents, typically by partitioning mentions into cross-document clusters. In the literature, the task has been formulated for entities, events, mixed entity–event settings, cross-domain data, hierarchical scientific concepts, and large-scale retrieval scenarios. At the same time, the field has remained fragmented by heterogeneous dataset formats, varying annotation standards, and the predominance of event coreference resolution as the de facto definition of CDCR, especially through the central role of ECB+ (Beheshti et al., 2013, Zhukova et al., 28 Feb 2026, Cattan et al., 2021).

1. Task scope and formal variants

In its classical form, CDCR seeks to determine when mentions in multiple documents refer to the same real-world referent and to cluster them accordingly. Early survey work described the task as identifying entity mentions across multiple documents that refer to the same underlying entity, emphasizing its role inside broader information extraction pipelines and its importance for large-scale document analysis (Beheshti et al., 2013). Subsequent work broadened the formulation to include events and concepts across many text documents, not only within a single domain or topic (Ravenscroft et al., 2021).

Modern work distinguishes several closely related formulations. Cross-document event coreference resolution (CDECR) clusters event mentions across documents that refer to the same real-world events (Min et al., 2024). Cross-document cross-domain coreference resolution (CD$^2$CR) links mentions across heterogeneous document types, specifically scientific work and newspaper articles that discuss them (Ravenscroft et al., 2021). Hierarchical CDCR (H-CDCR) extends ordinary clustering by jointly inferring both coreference clusters and hierarchy between them for scientific concepts (Cattan et al., 2021). Cross-document coreference search reformulates the task as retrieval: given a query event mention in context, the system must find all coreferring mentions in a large document collection (Eirew et al., 2022).

This range of formulations suggests that CDCR is not a single narrowly delimited benchmark problem. A plausible implication is that model design and evaluation criteria depend strongly on whether the target setting privileges strict identity, near-identity, hierarchy, domain transfer, or large-scale retrieval.

2. Annotation schemes and dataset ecology

A central distinction in CDCR research is the annotation scheme. ECB+ is an event-centric CDCR dataset with a strict annotation scheme focusing on identity coreference relations. It annotates mentions that describe participants, actions, times, and locations of events; pronouns can be included; and its span policy is mixed, with locations and times using maximum span while participants use minimum span (Zhukova et al., 2021). NewsWCL50 is concept-centric and annotates mentions based on frequently reported concepts in news articles, irrespective of whether they are strictly event-centric. It includes a mix of strict identity, loose, and bridging coreference relations, does not include pronouns, and uses maximum span throughout (Zhukova et al., 2021).

The comparison between ECB+ and NewsWCL50 established two recurring dimensions of dataset difficulty: lexical disambiguation and lexical diversity. The paper reports that NewsWCL50 has average chain size $35.0$, average unique lemmas per chain $10.5$, phrasing diversity $9.7$, inter-coder reliability $0.65$ (AOA), and coreference F1 without singletons $46.5$; ECB+ has average chain size $6.4$, average unique lemmas per chain $2.4$, phrasing diversity $1.4$, inter-coder reliability $0.76$ (Kappa), and coreference F1 without singletons $35.0$0 (Zhukova et al., 2021). The same work introduced phrasing diversity (PD) as a chain-level metric: $35.0$1 where $35.0$2 is the set of unique heads in chain $35.0$3, $35.0$4 is the number of unique full phrases with head $35.0$5, and $35.0$6 is the total number of phrases with head $35.0$7 (Zhukova et al., 2021).

Later work revised these schemes rather than merely contrasting them. A revised annotation of NewsWCL50 treats coreference chains as discourse elements (DEs), defined as language-independent semantic units that may be entities, events, groups, countries, organizations, or abstract concepts. The scheme links mentions by both identity and near-identity relations, including meronymic and metonymic relations, euphemism, metaphor, evaluative labeling, copular constructions, and aggregating relations (Zhukova et al., 19 Feb 2026). The reannotated test portions report the following statistics: NewsWCL50 has $35.0$8 chains, average chain size $35.0$9, UL $10.5$0, PD $10.5$1, MTLD $10.5$2, and CoNLL F1 $10.5$3; NewsWCL50$10.5$4 has $10.5$5 chains, average chain size $10.5$6, UL $10.5$7, PD $10.5$8, MTLD $10.5$9, and CoNLL F1 $9.7$0; ECB+ has $9.7$1 chains, average chain size $9.7$2, UL $9.7$3, PD $9.7$4, MTLD $9.7$5, and CoNLL F1 $9.7$6; ECB+$9.7$7 has $9.7$8 chains, average chain size $9.7$9, UL $0.65$0, PD $0.65$1, MTLD $0.65$2, and CoNLL F1 $0.65$3 (Zhukova et al., 19 Feb 2026).

These findings formalized a major controversy in CDCR annotation. Stricter identity-only schemes yield higher agreement and often higher benchmark scores, while looser, discourse-aware schemes better reflect varied and polarized reporting but reduce annotator agreement and expose models to harder semantic decisions. This suggests that dataset choice is inseparable from the operational definition of “coreference” itself.

3. Evaluation protocols and dataset unification

A major line of work argues that CDCR results were historically inflated by lenient or inconsistent evaluation. A pragmatic evaluation methodology was proposed that assumes access only to raw text rather than gold mentions, disregards singleton prediction, and evaluates at topic level or corpus level rather than on gold subtopics (Cattan et al., 2020). In that study, including singletons raised $0.65$4 F1 and CEAFe F1 for events by $0.65$5 and $0.65$6 points, respectively, and a system predicting all singletons could still obtain non-trivial scores if singletons were included (Cattan et al., 2020). Under earlier settings the authors report event performance of $0.65$7 F1, whereas under the more realistic topic-level predicted-mention setting performance drops to $0.65$8 F1 (Cattan et al., 2020).

The first end-to-end model for CDCR from raw text then established baseline results over predicted mentions on ECB+, reporting CoNLL F1 $0.65$9 for events, $46.5$0 for entities, and $46.5$1 for the unified ALL setting (Cattan et al., 2021). This made explicit that mention detection and cross-document clustering must be evaluated jointly if the goal is deployable CDCR rather than oracle-span clustering.

A complementary critique concerns corpus specificity. A uniform evaluation across ECB+, GVC, and FCC-T showed that systems trained on a single corpus do not generalize reliably to others, that document preclustering can artificially help on ECB+ but severely hurt on FCC-T, and that evaluation on multiple CDCR corpora is strongly necessary (Bugert et al., 2020). In the same direction, a discourse-coherence-based model trained on all corpora at once improved average performance across all datasets by $46.5$2 F1 points, while only modestly sacrificing per-corpus state of the art (Held et al., 2021).

The most systematic response to fragmentation is uCDCR, a unified benchmark that consolidates $46.5$3 publicly available English CDCR datasets into a consistent format, includes both entity and event coreference, corrects known inconsistencies, and enriches datasets with missing attributes (Zhukova et al., 28 Feb 2026). Its analysis reports that ECB+ has one of the lowest lexical diversities, that its CDCR complexity under the same-head-lemma baseline lies in the middle among all uCDCR datasets, and that the almost identical performance of the same-head-lemma baseline on events and entities shows that resolving both types is a complex task and should not be steered toward event coreference resolution alone (Zhukova et al., 28 Feb 2026).

4. Modeling strategies

Recent CDCR models differ primarily in how they constrain candidate pairs, represent mention context, and exploit higher-order or external information. End-to-end neural models adapted from within-document coreference resolution score spans and mention pairs across documents and then apply agglomerative clustering (Cattan et al., 2021, Cattan et al., 2020). One such model uses span representations

$46.5$4

and pair scores

$46.5$5

with $46.5$6 computed from $46.5$7 (Cattan et al., 2021). A related end-to-end model proposed a standardized evaluation protocol and reported ablation losses of $46.5$8 F1 without mention-scorer pre-training, $46.5$9 without dynamic pruning, $6.4$0 when switching from RoBERTa to BERT Large, and $6.4$1 without negative sampling (Cattan et al., 2020).

Sequential higher-order modeling offers a different approximation. A sequential cross-document model incrementally composes mentions into cluster representations and predicts links between a mention and the already constructed clusters, thereby approximating higher-order inference (Allaway et al., 2021). On ECB+ it reports $6.4$2 CoNLL F1 for entity coreference and $6.4$3 for event coreference, with incremental cluster composition contributing $6.4$4 F1 for entities and $6.4$5 for events over the corresponding cross-document model without candidate composition (Allaway et al., 2021).

Candidate pruning can also be learned from discourse. A two-stage system inspired by discourse coherence theory models reader focus as a $6.4$6-nearest-neighbor neighborhood in a learned latent embedding space and then applies a cross-encoder only to hard positives and negatives inside those neighborhoods (Held et al., 2021). It reports $6.4$7 F1 of $6.4$8 on ECB+ event coreference, $6.4$9 on GVC, $2.4$0 on FCC, $2.4$1 on ECB+ entity coreference, and $2.4$2 on CD2CR entity coreference (Held et al., 2021).

Large-language-model augmentation has become prominent, but usually as supervision or representation enrichment rather than direct clustering. A collaborative LLM–SLM approach first prompts an LLM to elaborate event mentions in context and then fine-tunes an SLM on fused original and summarized representations, achieving CoNLL F1 $2.4$3 on ECB+, $2.4$4 on GVC, and $2.4$5 on FCC, while GPT-4 alone attains $2.4$6, $2.4$7, and $2.4$8 on the same datasets (Min et al., 2024). A related approach uses LLM-generated free-text rationales as distant supervision and knowledge distillation, yielding $2.4$9 F1 $1.4$0 on ECB+, $1.4$1 on GVC, and $1.4$2 on AIDA Phase 1 (Nath et al., 2024).

Other work focuses on discourse structure, semantics, and debiasing. DIE-EC constructs document-level RST trees and cross-document lexical chains, merges them into a heterogeneous graph, and applies a GAT before pairwise scoring and agglomerative clustering (Gao et al., 2024). On WEC-Eng it reports CoNLL F1 $1.4$3 versus $1.4$4 for the WEC-Eng baseline; on WEC-Zh, $1.4$5 versus $1.4$6; and on ECB+, $1.4$7 (Gao et al., 2024). ACCI instead targets trigger bias in cross-document event coreference through a structural causal graph, backdoor-adjusted intervention,

$1.4$8

a counterfactual module, and an argument-aware enhancement module; it reports CoNLL F1 $1.4$9 on ECB+ and $0.76$0 on GVC (Yao et al., 2 Jun 2025).

For semantically loose and lexically rich mentions, unsupervised sieves remain competitive. XCoref applies a five-stage sieve architecture over named entities, non-named entities, groups of persons, and events or abstract entities, and on NewsWCL50 reports MUC F1 $0.76$1, $0.76$2 F1 $0.76$3, CEAF$0.76$4 F1 $0.76$5, and CoNLL F1 $0.76$6, outperforming both a lemma baseline and prior systems evaluated there (Zhukova et al., 2021).

5. Domain-specific extensions and alternative settings

The assumption that CDCR is mainly news event clustering has been challenged repeatedly. CD$0.76$7CR introduces a cross-domain setting linking entities across scientific articles and news articles. Its English dataset contains $0.76$8 coreference pair annotations over $0.76$9 documents, split into train $35.0$00 documents with $35.0$01 mentions and $35.0$02 clusters, dev $35.0$03 documents with $35.0$04 mentions and $35.0$05 clusters, and test $35.0$06 documents with $35.0$07 mentions and $35.0$08 clusters; mean Fleiss’ Kappa is $35.0$09, with $35.0$10 on the difficult subset (Ravenscroft et al., 2021). In this setting, a RoBERTa-based baseline trained directly on CD$35.0$11CR performs better than models transferred from ECB+, indicating that cross-domain generalization cannot be assumed (Ravenscroft et al., 2021).

Scientific literature motivates even stronger departures from ordinary CDCR. SciCo defines hierarchical CDCR for scientific concepts and provides an expert-annotated dataset with more than $35.0$12 mentions from more than $35.0$13 scientific documents, about $35.0$14 clusters, and about $35.0$15 hierarchical relations (Cattan et al., 2021). A unified Longformer-based joint model reports coreference CoNLL F1 $35.0$16, hierarchy F1 $35.0$17, and path ratio $35.0$18, compared with $35.0$19, $35.0$20, and $35.0$21 for the best baseline listed there (Cattan et al., 2021).

Wikipedia has been used to construct large-scale event-centric resources that remove topic restrictions. WEC-Eng contains $35.0$22 train mentions in $35.0$23 clusters, $35.0$24 dev mentions in $35.0$25 clusters, and $35.0$26 test mentions in $35.0$27 clusters; the corresponding CoNLL F1 on test is $35.0$28 for the proposed baseline versus $35.0$29 for a lemma baseline (Eirew et al., 2021). A search-oriented derivative, CoreSearch, contains more than $35.0$30 million passages, $35.0$31 validated non-singleton event clusters, and $35.0$32 mentions in dev and test, and uses a DPR-style retriever plus an integrated reader with coreference-based passage selection (Eirew et al., 2022).

Multilingual and domain-specific settings have also emerged. WEC-Zh provides a large-scale Chinese cross-document event coreference dataset with $35.0$33 event mentions and $35.0$34 clusters, developed because existing cross-document event coreference datasets were limited to English (Gao et al., 2024). In scientific software mention coreference, a hybrid system combining Sentence-BERT embeddings, FAISS-based centroid lookup, and HDBSCAN reports CoNLL F1 $35.0$35, $35.0$36, and $35.0$37 on Shared Task subtasks $35.0$38, $35.0$39, and $35.0$40, respectively (Matela et al., 25 Mar 2026). A knowledge-graph-oriented formulation further links textual mentions to graph entities using contextual embeddings, dynamic linking, and graph-based inference, and reports benchmark results such as $35.0$41 F1 for Llama-3 on CoNLL-2012 and $35.0$42 for ThaiCoref on the same dataset (Dong et al., 8 Apr 2025).

These extensions show that CDCR is increasingly treated as a family of relation induction problems rather than a single benchmark template. A plausible implication is that future progress will be constrained less by model capacity alone than by the match between annotation target, domain ontology, and inference regime.

6. Persistent challenges and research directions

Three challenges recur across the literature. The first is lexical and semantic variation. NewsWCL50 and its revisions show that near-identity, framing, metaphor, metonymy, and evaluative labeling are ordinary rather than exceptional in political news (Zhukova et al., 2021, Zhukova et al., 19 Feb 2026). SciCo shows that scientific concepts can be both lexically diverse and hierarchically structured (Cattan et al., 2021). CD$35.0$43CR shows that even apparently straightforward entity links become difficult when one side is scientific prose and the other is journalistic paraphrase (Ravenscroft et al., 2021).

The second challenge is generalizability. Cross-corpus analysis showed that systems developed on one corpus are often “hit-and-miss” on another and that models overfit on the structure of ECB+ (Bugert et al., 2020). Dataset unification work similarly argues that ECB+ covers only part of the mention and document diversity found across public CDCR resources, and that training and evaluation on all uCDCR datasets will improve the generalizability of CDCR models (Zhukova et al., 28 Feb 2026). Multi-corpus training and multi-scheme evaluation have therefore become explicit recommendations rather than optional robustness checks (Held et al., 2021, Zhukova et al., 2021).

The third challenge is evaluation realism. Work on streamlined evaluation argues for raw-text input, exclusion of singletons from scoring, and topic- or corpus-level evaluation without gold subtopic shortcuts (Cattan et al., 2020). Related work recommends reporting LEA alongside CoNLL F1, and reporting results with and without document preclustering because preclustering can mask weaknesses in the coreference model itself (Bugert et al., 2020).

Accordingly, current research directions converge on broader benchmarks, more explicit semantics, and stronger discourse modeling. The authors of the diverse-annotation comparison propose combining CDCR datasets with multiple annotation schemes that focus on various properties of the coreference chains (Zhukova et al., 2021). The revised NewsWCL50 and ECB+ annotations aim at balanced, discourse-aware CDCR in the news domain (Zhukova et al., 19 Feb 2026). Search-based formulations point toward large collections and targeted retrieval (Eirew et al., 2022). LLM-assisted systems suggest that generated summaries or rationales can inject contextual or abductive knowledge into smaller supervised models without replacing them (Min et al., 2024, Nath et al., 2024). Causal approaches indicate that some of the remaining error budget is due not only to missing information but also to confounding reliance on triggers rather than arguments (Yao et al., 2 Jun 2025).

Taken together, these developments position CDCR as a technically heterogeneous area spanning clustering, retrieval, discourse modeling, causal debiasing, and dataset design. The field’s central unresolved question is no longer merely how to cluster mentions across documents, but which cross-document referential relations should be modeled, under which annotation assumptions, and with which evaluation regime.