Joint Extraction of Temporal and Causal Event Relations

This direction is merged into another direction: temporal-causal integration. Its last version stays readable and citable here.

Version 1 of 111 papers17 conceptsfrozen 2026-10-04

When we read a story, we naturally weave events into a coherent tapestry. If we read that "the heavy rains caused the dam to burst," we don't just identify two events; we instantly understand that the rain happened first (temporal) and that it was the reason for the burst (causal).

In traditional Natural Language Processing, however, these dimensions are often treated as isolated silos. One model determines "when" things happened, while another determines "why." This fragmentation ignores a fundamental truth: time and causality are governed by strict physical and logical constraints. Recent research has shifted toward joint event relation extraction, a paradigm that treats these relations as interdependent to build more accurate and logically consistent models of the world.

The Logic of Physical Constraints

The core intuition behind this research is that temporal and causal relations are not independent variables. They are bound by a physical reality: a cause must precede or coincide with its effect.

As Paramita Mirza argues in Extracting Temporal and Causal Relations between Events, treating these as separate tasks prevents models from using temporal order to validate causal links. If a model predicts that Event A caused Event B, but also predicts that Event B happened before Event A, it has created a logical impossibility.

By modeling these relations jointly, researchers like Ning Qiang and Dan Roth create a bidirectional feedback loop. In Joint Reasoning for Temporal and Causal Relations, they demonstrate that causal "anchors" can resolve temporal ambiguities, while established temporal orders can filter out impossible causal candidates.

The complexity of event structures extends beyond simple A-to-B links. Events often exist in hierarchies where one event is a sub-component of another (e.g., "the surgery" includes "the incision").

Haoyu Wang and Muhao Chen expanded the joint framework to include these hierarchical relations. In Joint Constrained Learning for Event-Event Relation Extraction, they show that subevents are logically constrained to occur during their parent events. By converting these logical rules into differentiable loss functions, they allow models to learn from the inherent structure of events even when explicit labels for every relation type are missing.

This unified view is further supported by the creation of massive datasets like MAVEN-ERE. Developed by Xiaozhi Wang and colleagues, MAVEN-ERE provides a million-scale benchmark covering coreference, temporal, causal, and subevent relations, finally giving data-hungry neural models the breadth of information needed to see the "big picture" of event interactions.

From Classification to Structural Reasoning

Modern approaches are moving away from "black-box" classification—where a model simply guesses a label for a pair of events—toward frameworks that reason about the nature of the relations themselves.

  1. Prototype Matching: Zhilei Hu and Zixuan Li proposed ProtoEM, a framework that learns representative "prototypes" for different relation types. As detailed in ProtoEM: A Prototype-Enhanced Matching Framework for Event Relation Extraction, this allows the model to better handle data-scarce scenarios by understanding that a "cause" prototype inherently shares features with a "temporal before" prototype.
  2. Code-as-Prompt: To handle the messiness of long documents, Xingwei Tan and Yulan He developed CALLMSAE. In Cascading Large Language Models for Salient Event Graph Generation, they use Large Language Models to generate "salient event graphs." By prioritizing important events and using structured code prompts, they ensure that the resulting narrative is not just a list of facts, but a logically sound graph where hierarchy, time, and cause all align.

Why It Works

The success of joint extraction lies in constraint satisfaction. When a model is forced to satisfy multiple rules at once—transitivity (if A < B and B < C, then A < C), parent-child containment, and causal-temporal precedence—the search space for possible interpretations of a text shrinks. This reduction in ambiguity leads to "global" consistency, ensuring that the extracted event map isn't just a collection of local guesses, but a coherent reflection of the narrative's logic.

Papers in this version

History & citationVersion 1 · 2026-10-04

Merged into another direction · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_joint_extraction_of_temporal_and_causal__431f8e42,
  title = {Joint Extraction of Temporal and Causal Event Relations},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/joint-extraction-of-temporal-and-causal-event-relations/rdr_431f8e421ba14dae8f52d18487cc4185},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 1, frozen text):

@misc{lacuna_joint_extraction_of_temporal_and_causal__431f8e42_v1,
  title = {Joint Extraction of Temporal and Causal Event Relations},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/joint-extraction-of-temporal-and-causal-event-relations/rdr_431f8e421ba14dae8f52d18487cc4185/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}