Joint Coreference Resolution for Entities and Events

44 concepts30 papers1 proposal

When humans read a narrative, we naturally connect the actors (entities) with the actions they perform (events). If a news report mentions that "Alphabet acquired a quantum startup" and a later paragraph discusses "the $1B purchase," we instantly understand two things: "Alphabet" is the buyer, and "acquired" and "purchase" refer to the exact same real-world event.

In classical Natural Language Processing (NLP), however, these tasks have long been isolated. Systems typically run independent pipelines: one to group entity mentions (entity coreference), another to group event triggers (event coreference), and a third to map those mentions to a knowledge base (entity linking).

This isolated approach suffers from a fundamental flaw: error propagation. If an early pipeline stage makes a mistake due to weak local context, subsequent stages are forced to build on top of that error. Joint coreference modeling solves this by treating entities, events, and their links as a single, interconnected web. By allowing information to flow bi-directionally, high-confidence decisions in one task can resolve deep ambiguities in another.


The Symbiosis of Actions and Participants

The core intuition behind joint entity-event coreference is simple: identical events typically involve identical participants. If two sentences describe an action using different verbs (e.g., "the company bought the patent" vs. "the acquisition occurred on Tuesday"), identifying that "the company" and "the buyer" refer to the same entity provides a massive statistical signal that "bought" and "acquisition" are the same event.

[Entity Cluster 1] <---> [Event Cluster] <---> [Entity Cluster 2]
   "Alphabet"               "acquired"            "a startup"
       ^                        ^                      ^
       |                        |                      |
 "The tech giant" <--->  "the purchase"  <--->   "the company"

Early work by Heeyoung Lee et al. in their paper Joint Entity and Event Coreference Resolution across Documents operationalized this via an iterative clustering framework. Instead of making a single pass, the system allows semantic role dependencies to act as bridges. Information about event participants is passed to the entity resolver, which updates its clusters; these updated entity clusters are then fed back to refine the event clusters.

Shany Barhom et al. modernized this concept in Revisiting Joint Modeling of Cross-document Entity and Event Coreference Resolution by implementing a neural architecture that alternates between entity and event clustering. By using the cluster assignments of arguments as dynamic features for event matching, the model can resolve highly ambiguous verbs that would look entirely unrelated to a standalone lexical model. As Chris Tanner argues in Cross-Document Coreference Resolution for Entities and Events, this creates a symbiotic reinforcement loop: knowing the actors are identical increases the probability of matching the action, and matching the action locks in the identity of the actors.

To handle these interactions globally across multiple documents, Duy Phung et al. introduced message-passing via Hierarchical Graph Convolutional Networks for Jointly Resolving Cross-document Coreference of Entity and Event Mentions. By constructing sentence-level dependency trees and document-level interaction graphs, their model allows structural constraints to propagate across the entire corpus, ensuring that entity and event decisions remain globally consistent.


Propagating Knowledge via Entity Linking

A similar synergy exists between coreference resolution (grouping mentions together) and entity linking (grounding those mentions to a unique ID in a knowledge base like Wikidata).

Consider a multiparty conversation where speakers frequently use ambiguous pronouns like "he" or "she". A standalone entity linker cannot ground a isolated pronoun. However, if a joint model can establish a coreference link between that pronoun and a fully specified proper noun elsewhere in the text, the entity linker can easily ground the entire chain.

"He won the election."  <-- (Ambiguous pronoun: hard to link)
         ^
         |  [Coreference Link]
         v
"Barack Obama took the stage."  --> [Entity Link] --> KB: Q76 (Barack Obama)

In C3EL: A Joint Model for Cross-Document Co-Reference Resolution and Entity Linking, Sourav Dutta 0001 et al. demonstrated that this relationship is a two-way street. While coreference provides the global context needed for disambiguation, the structured features from a knowledge base (such as entity types or canonical aliases) provide strong anchor points that prevent the coreference system from merging unrelated clusters.

This mutual information exchange is often implemented via shared loss functions. For instance, Jiaxin Bai et al. developed a model for multiparty dialogues in Joint Coreference Resolution and Character Linking for Multiparty Conversation. Their architecture allows high-confidence signals from named entity linking to propagate backward through a shared optimization space, automatically correcting mistaken pronoun resolutions in the coreference chain.

Taking this a step further, Dhruv Agarwal et al. proposed in Entity Linking via Explicit Mention-Mention Coreference Modeling that entity linking should be framed directly as a joint clustering problem. By training dual-encoders to construct directed trees (arborescences) that connect mentions both to each other and to knowledge base entries simultaneously, the model leverages multiple surface forms in the text to build a more robust, noise-resistant representation of the underlying entity.


Unified Architectures and Soft Constraints

Instead of running separate models that talk to each other iteratively, can we collapse these tasks into a single, unified neural representation?

Vincent Ng 0001 explored this in a series of papers, including Constrained Multi-Task Learning for Event Coreference Resolution and Span-Based Event Coreference Resolution. Rather than treating entity and event coreference as distinct problems, Ng designed a unified span-based architecture that shares parameters across both mention types. To prevent the model from making logically impossible decisions, the framework enforces "soft" consistency constraints via penalty functions during training. This forces the network to learn an underlying representation of language where actions and participants are intrinsically bound together.

A similar philosophy drives joint information extraction. In Extracting Entities and Events as a Single Task Using a Transition-Based Neural Model, Junchi Zhang et al. utilized an incremental, state-transition process to build a joint entity-event structure from left to right. Their work proves that incorporating event-level context directly constrains and enhances the accuracy of entity detection, showing that we cannot fully know what an entity is without simultaneously understanding the role it plays in the surrounding events.

Key Foundational Work

An Entity-Level Approach to Information Extraction (2010) — contributed 1 concept elements to this direction.
Graph-Cut-Based Anaphoricity Determination for Coreference Resolution (2009) — contributed 1 concept elements to this direction.
Recognizing linked events: Searching the space of feasible explanations (2009) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_joint_coreference_and_entity_linking_3f5cfd27,
  title = {joint coreference and entity linking},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/joint-coreference-and-entity-linking/rdr_3f5cfd27375f473aad3f66690cf513bb},
  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_joint_coreference_and_entity_linking_3f5cfd27_v1,
  title = {joint coreference and entity linking},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/joint-coreference-and-entity-linking/rdr_3f5cfd27375f473aad3f66690cf513bb/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake