Global Coherence for Document Level Entity Resolution

You are reading version 1 of this direction, frozen 2026-10-04. The latest version is version 2.

Version 1 of 2221 papers276 conceptsfrozen 2026-10-04

In natural language processing, identifying an entity like "Ford" in a single sentence is often an impossible task. Is it the motor company, the 38th U.S. President, or a shallow place to cross a river? When a model looks only at a narrow window of text—the "local context"—it lacks the evidence to decide.

The research direction of Global Coherence shifts the focus from individual mentions to the entire document. By treating a document as a semantically consistent unit, researchers have developed mechanisms to ensure that all entities within a text "make sense" together.

The Failure of Local Windows

Traditional Named Entity Recognition (NER) and Disambiguation (NED) systems often process text sentence-by-sentence. As Arzoo Katiyar and Claire Cardie argue in Nested Named Entity Recognition Revisited, this localized approach limits a model's ability to resolve ambiguities that require broader textual evidence.

In long-form narratives, the crucial clue for an entity's identity might be established hundreds of tokens away. Arthur Amalvy and colleagues demonstrated in The Role of Global and Local Context in Named Entity Recognition that retrieving relevant sentences from anywhere in a document is often more effective than simply expanding the local neighborhood. Without this "global view," models frequently produce inconsistent labels for the same entity across different sections of the same text.

Collective Disambiguation as a Global Clique

One of the most powerful intuitions in this field is that entities in a document are not independent; they are topically related. Sayali Kulkarni, Soumen Chakrabarti, and their team formalized this in Collective annotation of Wikipedia entities in web text.

They framed entity disambiguation as a collective optimization problem. Instead of picking the best candidate for "Ford" and then the best for "Detroit" independently, the system treats the document as a single clique in a graphical model. It balances two signals:

  1. Local Compatibility: How well does the entity match the surrounding words?
  2. Global Coherence: How well do all selected entities relate to each other?

If the system identifies "General Motors" and "Detroit" elsewhere in the document, the global coherence signal pushes the ambiguous "Ford" toward the car manufacturer rather than the politician. This simultaneous disambiguation, as championed by Silviu Cucerzan in TAC Entity Linking by Performing Full-document Entity Extraction and Disambiguation, ensures the final set of entities is mutually consistent with the document's overall theme.

Joint Inference and the Coreference Bridge

A major breakthrough in achieving global consistency came from realizing that Entity Linking (EL) and Coreference Resolution are two sides of the same coin. Coreference identifies that "he," "the president," and "Gerald Ford" all refer to the same person; Entity Linking maps that person to a specific database entry.

In Joint Coreference Resolution and Named-Entity Linking with Multi-Pass Sieves, Hannaneh Hajishirzi and Luke Zettlemoyer showed that world knowledge can fix coreference errors, while document context can fix linking errors. If a coreference chain is established, the system can "pool" the context of every mention in that chain. Even if one mention is highly ambiguous, another mention in the same chain might be crystal clear.

This was taken further by Johannes Deleu and Chris Develder in Towards Consistent Document-level Entity Linking. They modeled the document as a directed tree where mentions and entities are nodes. This structure enforces a hard global constraint: all mentions within a coreferent cluster must link to the same entity. This prevents the logical inconsistency of a model identifying "Ford" as a company in paragraph one and a person in paragraph five.

Sequential Decisions and the Noise Challenge

While global coherence is a powerful signal, it can also introduce noise. If a document mentions "Ford" (the car) and "Lincoln" (the president), a naive global model might get confused.

To address this, researchers have moved toward more sophisticated decision-making processes. Joint Entity Linking with Deep Reinforcement Learning treats disambiguation as a sequence of decisions. The model starts with the "easiest" mentions—those with the strongest local cues—and uses those confirmed identities to build a reliable global context for the "harder," more ambiguous mentions.

Furthermore, Aixin Sun and Yi Tay argued in Pair-Linking for Collective Entity Disambiguation that forcing every entity to relate to every other entity (a dense clique) is often counterproductive. Instead, they suggest that semantic connections are sparse. Their "Pair-Linking" approach focuses on local clusters of related entities, capturing the "chain-like" nature of discourse without letting unrelated mentions distort the global signal.

By moving from isolated classification to collective, document-aware inference, these researchers have transformed entity resolution from a pattern-matching task into a holistic act of interpretation.

Papers in this version

Showing 60 of 221 papers.

History & citationVersion 2 · 2026-10-04

Current · Permanent link

  1. Version 2 (latest) · 2026-10-04 · 130 papers
  2. Version 1 · 2026-10-04 · 221 papers
Cite this direction

Latest version (a living page):

@misc{lacuna_global_document_coherence_e89a022d,
  title = {global document coherence},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/global-document-coherence/rdr_e89a022dab4d4e57991db61bee20bea2},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_global_coherence_for_document_level_enti_e89a022d_v1,
  title = {Global Coherence for Document Level Entity Resolution},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/global-coherence-for-document-level-entity-resolution/rdr_e89a022dab4d4e57991db61bee20bea2/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}