Cross-Document Entity Tracking — Cross-Document Coreference Resolution
Cross-Document Entity Coreference Resolution and Tracking · Also known as: cross-document coreference resolution, cross-doc entity linking, Belge Ötesi Varlık Takibi
Cross-document entity tracking, formally known as cross-document coreference resolution, identifies and merges all references to the same real-world entity scattered across a collection of documents. Rooted in the B3 evaluation framework introduced by Bagga and Baldwin (1998) and substantially advanced by the neural joint model of Barhom et al. (2019), the method builds entity clusters that span document boundaries — enabling multi-document understanding, knowledge-base population, and corpus-wide entity analysis.
Read the full method
Sign in with a free account to read this section.
Method map
The neighbourhood of related methods — select a node to explore.
When to use it
Cross-document entity tracking is appropriate when the analysis spans a defined collection of at least ten documents and requires knowing that a name in one document refers to the same entity as a different surface form in another. It is applied in knowledge-base construction (linking mentions to a unified entity record), multi-document summarisation (ensuring the same entity is counted once), news tracking (following a person or organisation across articles), and biomedical literature mining (consolidating gene or protein mentions). Two prerequisites must be satisfied before running this method: NER must have been completed across all documents, and the relevant document set must be explicitly defined. Without these, the method cannot begin.
Strengths & limitations
- Enables entity-level analysis across entire document collections — not just within individual documents.
- Produces structured entity clusters that serve directly as input to knowledge graphs, event timelines, and multi-document summaries.
- Neural joint models (Barhom et al., 2019) simultaneously resolve entity and event coreference, capturing richer relational context than earlier pipelined systems.
- The B3 metric (Bagga & Baldwin, 1998) provides a principled, mention-level evaluation that remains the standard for cross-document coreference benchmarks.
- Requires NER and within-document coreference resolution as completed upstream steps; errors in those steps propagate into the cross-document clusters.
- The document set must be defined before the analysis begins; the method does not self-select relevant documents.
- Evaluation requires gold-standard coreference annotations, which are expensive to produce for new domains or languages.
- Computational cost scales with the number of mention pairs across the collection; very large corpora need approximate indexing strategies.
Frequently asked
What is the difference between within-document and cross-document coreference?
Within-document coreference resolution groups mentions that refer to the same entity inside a single document — for example, linking 'Barack Obama', 'the President', and 'he' within one article. Cross-document coreference extends this to a collection of documents, linking mentions that appear in different documents but refer to the same real-world entity. The cross-document step depends on within-document resolution being completed first.
Why is B3 the standard metric rather than simple accuracy?
Simple accuracy does not handle the cluster structure of coreference well. B3, introduced by Bagga and Baldwin (1998), computes precision and recall for every individual mention by examining the overlap between the predicted cluster containing that mention and the gold cluster. This mention-level view is more informative than chain-level scoring and penalises both over-merging (putting too many mentions in one cluster) and over-splitting (creating too many small clusters).
Can the method work without gold-standard annotations?
It can produce clusters without annotations, but it cannot be rigorously evaluated. Without a gold standard, you can only inspect clusters manually or use indirect proxies such as internal cluster cohesion. For any research or production use where quality guarantees matter, a labelled reference set is necessary to measure B3, MUC, or CEAFe.
How large does the document collection need to be?
The statwise engine requires a minimum of ten documents for the clustering step to be meaningful. With very small collections the number of cross-document mention pairs is too low for cluster quality metrics to be stable. There is no strict upper limit, but very large collections require approximate nearest-neighbour indexing (such as FAISS) to keep computation tractable.
Sources
- Bagga, A. & Baldwin, B. (1998). Algorithms for Scoring Coreference Chains. In Proceedings of the LREC 1998 Linguistic Coreference Workshop, pp. 563–566. link ↗
- Barhom, S., Shwartz, V., Eirew, A., Bugert, M., Reimers, N. & Dagan, I. (2019). Revisiting Joint Modeling of Cross-document Entity and Event Coreference Resolution. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics (ACL), pp. 4179–4189. link ↗
How to cite this page
ScholarGate. (2026, June 1). Cross-Document Entity Coreference Resolution and Tracking. ScholarGate. https://scholargate.app/en/text-mining/cross-document-entity-tracking
Which method?
Set this method beside its closest kin and read them side by side — the library lays the books on the table; the choice is yours.
- Coreference ResolutionText mining↔ compare
- Entity LinkingText mining↔ compare
- Information ExtractionText mining↔ compare
- Named Entity RecognitionText mining↔ compare