Document Level Temporal Relation Extraction

147 concepts91 papers1 proposal

Imagine reading a patient's clinical history. The very first line states: "Patient admitted on October 12." Several paragraphs down, the note mentions: "The patient developed a high fever after starting the new medication."

To understand the chronological order of these events, a human reader effortlessly connects the fever back to the admission date. For a machine, however, this is exceptionally difficult. Traditional Natural Language Processing (NLP) models are built to look at local context—usually a single sentence or a narrow window of adjacent sentences. When critical temporal anchors and events are separated by paragraphs of text, these local models suffer from a form of amnesia.

Document-level temporal relation extraction is the field dedicated to solving this problem: teaching machines to build a globally coherent timeline across an entire text, rather than just parsing time within isolated clauses.

The Local Context Bottleneck

Why do standard NLP models fail when moving from single sentences to full documents?

As Alfio Gliozzo et al. demonstrate in [When Did that Happen?

  • Linking Events and Relations to Timestamps](/work/when-did-that-happen-linking-events-and-relations), a substantial portion of temporal information is inherently distributed across sentence boundaries. When models are restricted to localized annotation windows, they fail to capture these long-distance dependencies, resulting in event graphs that lack global completeness (Taylor Cassidy et al. in An Annotation Framework for Dense Event Ordering).

If a model only understands local context, the naive way to build a document timeline is to evaluate every possible pair of events and time expressions. However, this creates two massive bottlenecks:

  1. The Quadratic Explosion: If a document has $N$ events, a pairwise model must classify $O(N^2)$ relationships. As SHUAICHENG ZHANG et al. point out in Extracting Temporal Event Relation with Syntax-guided Graph Transformer, scaling this to full documents becomes computationally intractable.
  2. Error Accumulation: To connect distant events, a local model must "chain" adjacent predictions (e.g., if Event A is before Event B, and Event B is before Event C, then Event A is before Event C). Artuur Leeuwenberg et al. observe in Temporal Information Extraction by Predicting Relative Time-lines that this chaining causes local errors to compound rapidly, leading to logically impossible timelines (such as cyclical loops where A happens before B, B before C, and C before A).

To extract long-range temporal dependencies accurately, researchers have shifted away from isolated local classifications toward mechanisms that enforce global coherence.

Enforcing Coherence via Global Constraints

One powerful way to resolve long-distance temporal relations is to force the model's predictions to obey the laws of formal logic across the entire document.

Instead of allowing a classifier to make independent guesses for every event pair, Quang Do et al. introduced a joint inference framework using Integer Linear Programming (ILP) in Joint Inference for Event Timeline Construction. By evaluating all pairs simultaneously and applying mathematical transitivity constraints (if $A \prec B$ and $B \prec C$, then $A \prec C$), the system uses the certain, local relations to automatically correct or infer the ambiguous, long-distance relations.

[Local Prediction]  Event A occurs BEFORE Event B  (High Confidence)
[Local Prediction]  Event B occurs BEFORE Event C  (High Confidence)
                       │
                       ▼
[Global Inference]  Event A must occur BEFORE Event C (Even if separated by paragraphs)

This global structural awareness does more than just clean up final predictions; it can also act as a teacher. Ning Qiang et al. demonstrated in Exploiting Partially Annotated Data in Temporal Relation Extraction that global transitivity constraints can be used to harvest "incidental supervision." By taking sparse, locally annotated data and propagating those labels across the document graph using transitivity, models can learn to resolve long-distance dependencies without requiring humans to manually label every single far-apart event pair.

Bypassing Pairs: Direct Timeline Projection

A more radical solution to the quadratic explosion and error accumulation problems is to abandon pairwise classification altogether.

In Temporal Information Extraction by Predicting Relative Time-lines, the authors proposed a model that maps events directly onto a continuous, linear relative timeline by predicting absolute start points and durations for each event. Because every event is placed on the same unified axis, logical consistency is mathematically guaranteed from the start, and the computational complexity drops from quadratic ($O(N^2)$) to linear ($O(N)$).

This concept of anchoring events to a shared narrative structure is highly effective in complex domains like medicine. As Guergana Savova et al. note in SemEval-2015 Task 6: Clinical TempEval, clinical notes rely heavily on "narrative containers"—broad time frames (like an admission period or a specific surgery day) that encompass multiple smaller events.

By training models to identify these overarching containers first, distant events can be anchored to the same temporal hub. Riza Batista-Navarro et al. leverage this in TIMELINE: Exhaustive Annotation of Temporal Relations Supporting the Automatic Ordering of Events in News Articles, showing that the overarching narrative structure of a document provides a much clearer chronological signal for distant events than local linguistic features provide for nearby ones.

Bridging Paragraphs with Discourse Graphs

How do modern deep learning architectures physically pass temporal context across multiple sentence boundaries? Local syntactic features, like dependency parse trees, naturally break down at sentence boundaries.

To bridge this gap, Puneet Mathur et al. developed TIMERS: Document-level Temporal Relation Extraction, which utilizes Gated Relational Graph Convolutional Networks (GR-GCNs). Instead of treating the document as a flat sequence of words, the model constructs a structural graph based on News Discourse Profiling.

Sentences are categorized by their functional role in the narrative—such as "Main Event," "Historical Context," or "Expectation." The graph network then passes information along these discourse edges.

[Sentence 1: Main Event] ───(Discourse Edge)───► [Sentence 4: Historical Context]
         │                                                 │
   (Local Event)                                     (Temporal Anchor)
         │                                                 │
         ▼                                                 ▼
   "Patient has fever"  ◄───────────────────────────  "Admitted on Oct 12"
                         (Long-Range Dependency Met)

By explicitly routing information through the discourse structure, the model can bypass the irrelevant text separating two sentences. This allows the network to perform multi-hop reasoning, linking an event in a "Main Event" sentence directly to a temporal anchor located paragraphs away in a "Historical Context" sentence.

Ultimately, document-level temporal extraction works because it stops treating documents as chains of words and starts treating them as interconnected webs of discourse, where global logic dictates local meaning.

History & citationVersion 2 · 2026-10-04

Current · Permanent link

  1. Version 2 (latest) · 2026-10-04 · 98 papers
  2. Version 1 · 2026-10-04 · 108 papers
Cite this direction

Latest version (a living page):

@misc{lacuna_long_range_temporal_dependency_long_dist_506f9861,
  title = {long-range temporal dependency (long-distance temporal dependency)},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/long-range-temporal-dependency-long-distance-temporal-dependency/rdr_506f986149fb4f2cb24211ab243b2438},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 2, frozen text):

@misc{lacuna_long_range_temporal_dependency_long_dist_506f9861_v2,
  title = {long-range temporal dependency (long-distance temporal dependency)},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/long-range-temporal-dependency-long-distance-temporal-dependency/rdr_506f986149fb4f2cb24211ab243b2438/v/2},
  note = {Version 2, frozen 2026-10-04. Accessed 2026-11-20},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake