---
title: Cross-Document Coreference Resolution
url: https://www.emergentmind.com/topics/cross-document-coreference-resolution-cdcr
type: topic
---

# Cross-Document Coreference Resolution

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+ [1311.3987, 2603.00621, 2104.08809].

## 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 [1311.3987]. Subsequent work broadened the formulation to include events and concepts across many text documents, not only within a single domain or topic [2101.12637].

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 [2406.02148]. Cross-document cross-domain coreference resolution (CD$^2$CR) links mentions across heterogeneous document types, specifically scientific work and newspaper articles that discuss them [2101.12637]. Hierarchical CDCR (H-CDCR) extends ordinary clustering by jointly inferring both coreference clusters and hierarchy between them for scientific concepts [2104.08809]. 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 [2210.12654].

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 [2109.05250]. 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 [2109.05250].

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 \(64.4\) [2109.05250]. The same work introduced phrasing diversity (PD) as a chain-level metric:
\[
PD_c = \sum_{h \in H_e} \frac{|U_h|}{|P_h|}
\]
where \(H_e\) is the set of unique heads in chain \(c\), \(|U_h|\) is the number of unique full phrases with head \(h\), and \(|P_h|\) is the total number of phrases with head \(h\) [2109.05250].

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 [2602.17424]. The reannotated test portions report the following statistics: NewsWCL50 has \(134\) chains, average chain size \(38.2\), UL \(10.71\), PD \(8.99\), MTLD \(14.54\), and CoNLL F1 \(48.05\); NewsWCL50\(_r\) has \(433\) chains, average chain size \(15.1\), UL \(6.05\), PD \(9.06\), MTLD \(15.58\), and CoNLL F1 \(54.08\); ECB+ has \(171\) chains, average chain size \(2.4\), UL \(2.19\), PD \(1.99\), MTLD \(4.55\), and CoNLL F1 \(69.91\); ECB+\(_r\) has \(97\) chains, average chain size \(14.7\), UL \(5.88\), PD \(9.04\), MTLD \(20.65\), and CoNLL F1 \(52.92\) [2602.17424].

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 [2009.11032]. In that study, including singletons raised \(B^3\) F1 and CEAFe F1 for events by \(+11.9\) and \(+17.8\) points, respectively, and a system predicting all singletons could still obtain non-trivial scores if singletons were included [2009.11032]. Under earlier settings the authors report event performance of \(71.1\) F1, whereas under the more realistic topic-level predicted-mention setting performance drops to \(48.2\) F1 [2009.11032].

The first end-to-end model for CDCR from raw text then established baseline results over predicted mentions on ECB+, reporting CoNLL F1 \(54.4\) for events, \(35.7\) for entities, and \(43.4\) for the unified ALL setting [2106.01210]. 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 [2011.12249]. In the same direction, a discourse-coherence-based model trained on all corpora at once improved average performance across all datasets by \(17.2\) F1 points, while only modestly sacrificing per-corpus state of the art [2110.05362].

The most systematic response to fragmentation is uCDCR, a unified benchmark that consolidates \(12\) publicly available English CDCR datasets into a consistent format, includes both entity and event coreference, corrects known inconsistencies, and enriches datasets with missing attributes [2603.00621]. 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 [2603.00621].

## 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 [2106.01210, 2009.11032]. One such model uses span representations
\[
g_i = [\text{XFIRST}(i),\, \text{XLAST}(i),\, \hat{x}_i,\, \phi(i)]
\]
and pair scores
\[
s(i,j) = s_m(i) + s_m(j) + s_a(i,j)
\]
with \(s_a(i,j)\) computed from \([g_i, g_j, g_i \circ g_j]\) [2106.01210]. A related end-to-end model proposed a standardized evaluation protocol and reported ablation losses of \(-3.2\) F1 without mention-scorer pre-training, \(-4.0\) without dynamic pruning, \(-4.1\) when switching from RoBERTa to BERT Large, and \(-1.4\) without negative sampling [2009.11032].

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 [2104.08413]. On ECB+ it reports \(75.3\) CoNLL F1 for entity coreference and \(80.8\) for event coreference, with incremental cluster composition contributing \(+8.9\) F1 for entities and \(+2.8\) for events over the corresponding cross-document model without candidate composition [2104.08413].

Candidate pruning can also be learned from discourse. A two-stage system inspired by discourse coherence theory models reader focus as a \(K\)-nearest-neighbor neighborhood in a learned latent embedding space and then applies a cross-encoder only to hard positives and negatives inside those neighborhoods [2110.05362]. It reports \(B^3\) F1 of \(86.6\) on ECB+ event coreference, \(83.0\) on GVC, \(63.5\) on FCC, \(82.8\) on ECB+ entity coreference, and \(78.5\) on CD2CR entity coreference [2110.05362].

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 \(86.7\) on ECB+, \(87.4\) on GVC, and \(78.7\) on FCC, while GPT-4 alone attains \(76.8\), \(10.2\), and \(6.1\) on the same datasets [2406.02148]. A related approach uses LLM-generated free-text rationales as distant supervision and knowledge distillation, yielding \(B^3\) F1 \(86.8\) on ECB+, \(84.3\) on GVC, and \(64.5\) on AIDA Phase 1 [2404.03196].

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 [2406.15990]. On WEC-Eng it reports CoNLL F1 \(65.0\) versus \(58.9\) for the WEC-Eng baseline; on WEC-Zh, \(66.9\) versus \(62.1\); and on ECB+, \(85.5\) [2406.15990]. ACCI instead targets trigger bias in cross-document event coreference through a structural causal graph, backdoor-adjusted intervention,
\[
P(Y \mid do(X)) = \sum_t P(Y \mid X, A, T=t) P(T=t),
\]
a counterfactual module, and an argument-aware enhancement module; it reports CoNLL F1 \(88.4\%\) on ECB+ and \(85.2\%\) on GVC [2506.01488].

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 \(84.7\), \(B^3\) F1 \(55.0\), CEAF\(_e\) F1 \(22.6\), and CoNLL F1 \(54.1\), outperforming both a lemma baseline and prior systems evaluated there [2109.05252].

## 5. Domain-specific extensions and alternative settings

The assumption that CDCR is mainly news event clustering has been challenged repeatedly. CD$^2$CR introduces a cross-domain setting linking entities across scientific articles and news articles. Its English dataset contains \(7{,}602\) coreference pair annotations over \(528\) documents, split into train \(300\) documents with \(4{,}604\) mentions and \(426\) clusters, dev \(142\) documents with \(1{,}821\) mentions and \(199\) clusters, and test \(86\) documents with \(1{,}177\) mentions and \(101\) clusters; mean Fleiss’ Kappa is \(0.554\), with \(0.399\) on the difficult subset [2101.12637]. In this setting, a RoBERTa-based baseline trained directly on CD$^2$CR performs better than models transferred from ECB+, indicating that cross-domain generalization cannot be assumed [2101.12637].

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 \(26{,}000\) mentions from more than \(20{,}000\) scientific documents, about \(10{,}000\) clusters, and about \(6{,}000\) hierarchical relations [2104.08809]. A unified Longformer-based joint model reports coreference CoNLL F1 \(77.2\), hierarchy F1 \(44.5\), and path ratio \(47.2\), compared with \(66.8\), \(23.8\), and \(28.4\) for the best baseline listed there [2104.08809].

Wikipedia has been used to construct large-scale event-centric resources that remove topic restrictions. WEC-Eng contains \(40{,}529\) train mentions in \(7{,}042\) clusters, \(1{,}250\) dev mentions in \(233\) clusters, and \(1{,}893\) test mentions in \(322\) clusters; the corresponding CoNLL F1 on test is \(62.3\) for the proposed baseline versus \(53.1\) for a lemma baseline [2104.05022]. A search-oriented derivative, CoreSearch, contains more than \(2.7\) million passages, \(522\) validated non-singleton event clusters, and \(2{,}893\) mentions in dev and test, and uses a DPR-style retriever plus an integrated reader with coreference-based passage selection [2210.12654].

Multilingual and domain-specific settings have also emerged. WEC-Zh provides a large-scale Chinese cross-document event coreference dataset with \(53{,}066\) event mentions and \(4{,}476\) clusters, developed because existing cross-document event coreference datasets were limited to English [2406.15990]. In scientific software mention coreference, a hybrid system combining Sentence-BERT embeddings, FAISS-based centroid lookup, and HDBSCAN reports CoNLL F1 \(0.98\), \(0.98\), and \(0.96\) on Shared Task subtasks \(1\), \(2\), and \(3\), respectively [2603.24246]. 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 \(73.9\) F1 for Llama-3 on CoNLL-2012 and \(78.7\) for ThaiCoref on the same dataset [2504.05767].

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 [2109.05250, 2602.17424]. SciCo shows that scientific concepts can be both lexically diverse and hierarchically structured [2104.08809]. CD$^2$CR shows that even apparently straightforward entity links become difficult when one side is scientific prose and the other is journalistic paraphrase [2101.12637].

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+ [2011.12249]. 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 [2603.00621]. Multi-corpus training and multi-scheme evaluation have therefore become explicit recommendations rather than optional robustness checks [2110.05362, 2109.05250].

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 [2009.11032]. 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 [2011.12249].

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 [2109.05250]. The revised NewsWCL50 and ECB+ annotations aim at balanced, discourse-aware CDCR in the news domain [2602.17424]. Search-based formulations point toward large collections and targeted retrieval [2210.12654]. LLM-assisted systems suggest that generated summaries or rationales can inject contextual or abductive knowledge into smaller supervised models without replacing them [2406.02148, 2404.03196]. 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 [2506.01488].

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.

Source: https://www.emergentmind.com/topics/cross-document-coreference-resolution-cdcr