Modeling Inter-Event and Intra-Event Dependencies

112 concepts80 papers2 proposals

Traditional event extraction systems often suffer from a form of "tunnel vision." They treat text like an assembly line of isolated classification tasks: first, identify a trigger word (e.g., "fired"); next, determine the event type (e.g., Employment-Termination or Attack); and finally, detect individual arguments (e.g., who was fired, or what weapon was used).

The flaw in this pipeline approach is that language is inherently relational. Event components do not exist in a vacuum. If a sentence mentions a "buyer," a "seller" is almost certainly present. If a document reports an "Attack," the probability of a "Die" or "Casualty" event nearby increases dramatically.

To exploit these natural patterns, a major research direction focuses on dependency modeling—explicitly capturing the semantic, statistical, and structural correlations both within a single event (intra-event dependencies) and across multiple events (inter-event dependencies). By shifting from isolated classification to joint, collective inference, these frameworks leverage global structural priors to resolve local ambiguities.

Intra-Event Dependencies: Modeling Inter-Argument Correlations

When extracting the arguments of an event, standard classification models often evaluate each candidate entity independently. This independence assumption fails to capture the logical constraints that bind participants together in real-world scenarios.

In Fast and Robust Joint Models for Biomedical Event Extraction, Sebastian Riedel et al. addressed this by moving away from pipeline systems toward joint models. Using dual decomposition, the authors enforced global consistency across complex multi-argument structures, allowing the system to learn which entities are likely to participate together in a shared context. Similarly, in RBPB: Regularization-Based Pattern Balancing Method for Event Extraction, Lei Sha et al. introduced a regularization framework to capture how the identification of one argument role significantly increases the likelihood of identifying related roles within the same event context.

Reframing the extraction mechanism can also make these dependencies explicit. In Bi-Directional Iterative Prompt-Tuning for Event Argument Extraction, Lu Dai proposed a cloze-style task that captures the "conversation" between different arguments. By predicting roles iteratively in both forward and backward directions, the model uses the identification of one role to constrain and inform the prediction of others.

To formalize these structures, Hao Li et al. developed the EDGE framework in Intra-Event and Inter-Event Dependency-Aware Graph Network for Event Argument Extraction. EDGE models events as structured templates with interdependent roles, allowing the system to leverage structural priors—such as automatically expecting a seller when a buyer is identified—to resolve local text ambiguities.

Inter-Event Dependencies: Cross-Event and Cross-Type Correlations

Events within a document are rarely isolated; they follow narrative schemas, causal chains, and temporal sequences. Traditional models that classify each trigger in isolation miss out on these macro-level patterns, struggling with polysemous words or implicit context.

In Using Document Level Cross-Event Inference to Improve Event Extraction, Shasha Liao et al. demonstrated that document-level cross-event relationships provide critical context clues. For instance, identifying a hiring event provides the thematic context needed to disambiguate subsequent mentions of employment termination from unrelated physical movements. To capture these links within sentences, Yubo Chen et al. introduced collective event detection in Collective Event Detection via a Hierarchical and Bias Tagging Networks with Gated Multi-level Attention Mechanisms, treating multiple occurrences within a single sentence as interdependent so that the classification of one trigger directly informs the interpretation of another.

This collective modeling becomes even more powerful when extended across sentence boundaries. In MLBiNet: A Cross-Sentence Collective Event Detection Network, Dongfang Lou et al. utilized a stacked architecture of bidirectional decoders to model the co-occurrence grammar of related events, such as linking a conflict event to subsequent casualties across a document.

Furthermore, the presence of multiple events can serve as a contrastive training signal. As shown by Yuxin He et al. in Revisiting Event Argument Extraction: Can EAE Models Learn Better When Being Aware of Event Co-occurrences?, exposing models to co-occurring events simultaneously prevents the loss of valuable boundary signals that occurs when multi-event sentences are decomposed into isolated training instances.

Unified Structural Frameworks for Joint Inference

To fully exploit both intra- and inter-event dependencies, modern architectures increasingly abandon local classifiers entirely, opting instead for global, structural prediction frameworks that parse text directly into unified graphs or energy spaces.

One prominent approach is to reframe event extraction as a graph parsing problem. In Event Detection as Graph Parsing, Jianye Xie used a pre-trained sequence-to-sequence model to generate linearized event graphs. This approach captures the global context and logical correlations between co-occurring event types in a single pass.

Similarly, joint modeling can resolve dependencies across different sub-tasks that traditionally conflict. In Span-Based Event Coreference Resolution, Vincent Ng 0001 introduced a unified span-based model that jointly learns trigger detection, argument extraction, and coreference resolution. By optimizing a unified loss function covering the entire event structure, the model exploits the mutual dependencies between all event elements simultaneously.

[Isolated Pipeline]
Trigger Detection ──> Role Assignment ──> Coreference Resolution (Errors cascade)

[Joint Dependency Framework]
┌────────────────────────────────────────────────────────┐
│  Trigger Type <───> Argument Roles <───> Coreference   │ (Globally optimized)
└────────────────────────────────────────────────────────┘

Alternatively, energy-based formulations offer a way to score the global coherence of an entire document's event structure. In SPEECH: Structured Prediction with Energy-Based Event-Centric Hyperspheres, Bryan Hooi developed a framework that uses Energy-Based Models (EBMs) to capture highly interdependent sub-tasks. Instead of making independent, greedy choices, the system evaluates the "energy" or structural consistency of the entire configuration, ensuring that local decisions align with global narrative constraints.

Key Foundational Work

Using Document Level Cross-Event Inference to Improve Event Extraction (2010) — contributed 3 concept elements to this direction.
Mining personal experiences and opinions from Web documents (2011) — contributed 2 concept elements to this direction.
Fast and Robust Joint Models for Biomedical Event Extraction (2011) — contributed 2 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_inter_event_dependency_modeling_6ca8f48a,
  title = {inter-event dependency modeling},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/inter-event-dependency-modeling/rdr_6ca8f48a1fca4227aa7130f4179bb6b5},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-21},
}

This exact version (version 1, frozen text):

@misc{lacuna_inter_event_dependency_modeling_6ca8f48a_v1,
  title = {inter-event dependency modeling},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/inter-event-dependency-modeling/rdr_6ca8f48a1fca4227aa7130f4179bb6b5/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-21},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake