Document Level Causal Reasoning

69 concepts55 papers1 proposal

When humans read a narrative, we effortlessly connect an event in the first paragraph to a consequence three paragraphs later. Traditional natural language processing models, however, are largely blind to this macro-logic. They excel at intra-sentence extraction—identifying that "heavy rain" caused a "flood" when both appear in the same clause. But as comprehensive surveys by Yang et al. and Cheng et al. point out, real-world causality is rarely confined to single sentences. It spans across sentence boundaries, paragraphs, and entire documents.

Document-level causal reasoning shifts the focus from local pattern matching to constructing a global mental model of a text's narrative logic.

Why Local Patterns Fail

Intra-sentence models rely heavily on explicit linguistic connectives like because, due to, or therefore. When causal dependencies span long ranges, these explicit markers often vanish. For instance, a corporate report might describe a strategic shift in section one and a profit increase in section four. There is no explicit connective linking them; the relationship must be inferred through global context and domain knowledge.

Traditional natural language processing tools fail here because they treat causal extraction as an isolated classification task over narrow windows of text. Furthermore, standard data augmentation techniques often destroy this delicate macro-logic by altering sentences without preserving the overarching causal chain, as noted by Zuo et al.. Historically, datasets have also truncated annotations to single sentences, leaving models starved of long-range training examples, a limitation that Tan et al. addressed by introducing the Causal News Corpus to bridge discourse-level and event-based data.

Graph-Based Structural Modeling

To capture long-range dependencies, a prominent strategy transforms linear text into structured relational graphs. Instead of reading tokens sequentially, models treat events as nodes and learn to predict the causal edges between them, bypassing the spatial separation that blinds standard architectures.

[Event A: Policy Change] ----(Syntax/Discourse Edge)----> [Context Clues]
           |                                                    |
           +-----------------(Causal Edge)---------------------> v
                                                      [Event B: Market Reaction]

Several distinct graph architectures have been proposed to execute this concept:

  • Integrating Multi-Source Knowledge: Phu (2021) developed RichGCN, a framework that builds document-level interaction graphs by incorporating syntax, discourse structures, and external lexical knowledge (like WordNet) to bridge distant sentences.
  • Contextual Transformers: To better connect events scattered throughout an extended text, Chen et al. introduced ERGO (Event Relational Graph Transformer), which leverages graph-based attention to track dependencies across sentence boundaries.
  • Filtering Noise via Interaction Graphs: Because a long document contains mostly unrelated noise, Pu et al. designed ECLEP. This framework first predicts which individual events are even capable of being "causality-related" before using a Graph Attention Network (GAT) to model how event pairs interact.
  • Hierarchical Stepping Stones: To make document-level graph reasoning tractable, Hu et al. proposed a progressive reasoning strategy (PPAT). It treats document-level causality as a hierarchical task: mastering local, explicit sentence-level causal cues provides the necessary foundation for inferring implicit, long-distance relations across the broader document.

Global Inference and Constraints

Local pairwise classification—deciding if event $A$ caused event $B$ by looking only at those two events—fails at the document level because it ignores the broader narrative flow. To overcome this, Gao et al. (2019) introduced a global inference framework using Integer Linear Programming (ILP).

Instead of evaluating event pairs in isolation, their approach optimizes for the entire document's narrative structure simultaneously. It models "causal structures"—the structural tendency for a story's main events to act as central hubs for causes and effects. By enforcing global structural constraints during inference, the model uses the document's overall topic transitions to resolve local ambiguities and uncover implicit, long-distance event pairs.

Generative and Cloze-Test Reframing

With the advent of large language models, researchers have increasingly shifted from discriminative classification to generative reasoning to capture deep semantic dependencies.

For example, Haoran Li (2024) reframed event causality identification as a "cloze test" via the Semantic Dependency Inquiry Network (SemDI). By masking one event and forcing the model to predict a fill-in token based on the surrounding context, the network is compelled to perform the multi-hop logical reasoning required to connect distant events within a single document.

Similarly, Liu et al. (2023) introduced the Implicit Cause-Effect interaction (ICE) framework, treating extraction as a generative task. They utilized a teacher-student distillation approach paired with an optimal transport alignment strategy to help the model move past simple pattern matching toward deep causal reasoning.

Taking generation a step further, Xiang et al. bypassed traditional classifiers entirely. They instructed language models to generate structured "causality statements" and then evaluated the mathematical log-probabilities of those statements to determine the true direction of the causal flow.

Multi-Step Chains and Cross-Document Synthesis

Real-world document reasoning ultimately requires tracking nested causal chains, where an initial effect acts as the trigger for a subsequent event. As Moghimifar et al. argue, mapping these multi-stage relationships is essential for technical, medical, or financial documentation.

This tracking manifests across several specialized domains:

  • Scientific Discovery: In the biomedical domain, Surdeanu et al. and Hope et al. focused on extracting "causal precedence" and mechanisms, piecing together fragmented interactions across full-text documents into coherent biological signaling pathways.
  • Financial Explanations: Kang Liu (2021) introduced FinReason to link structured database records (such as corporate performance changes) to their narrative justifications buried deep within long financial reports.
  • Counterfactual Generation: Wang et al. integrated Structural Causal Models (SCMs) with disentangled representation learning to ensure that when a narrative context changes, the counterfactual updates logically cascade through long-form documents.

The ultimate frontier of this research direction pushes beyond the boundaries of a single text. Romanou et al. introduced the CRAB benchmark to evaluate cross-document event pairing, where causal links must be inferred between events described in entirely separate news sources. Their findings highlight the core challenge of the field: while models are proficient at identifying local linguistic cues like because, their performance drops significantly when forced to synthesize information across documents. This underscores the ongoing shift toward building systems that construct true, document-agnostic mental models of cause and effect.

Key Foundational Work

Emotion Cause Detection with Linguistic Constructions (2010) — contributed 2 concept elements to this direction.
Learning Strategies for Open-Domain Natural Language Question Answering (2005) — contributed 1 concept elements to this direction.
History & citationVersion 2 · 2026-10-04

Current · Permanent link

  1. Version 2 (latest) · 2026-10-04 · 55 papers
  2. Version 1 · 2026-10-04 · 76 papers
Cite this direction

Latest version (a living page):

@misc{lacuna_document_level_causal_reasoning_2117217c,
  title = {document-level causal reasoning},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/document-level-causal-reasoning/rdr_2117217c43d342dcbaaad19f7325c4b6},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 2, frozen text):

@misc{lacuna_document_level_causal_reasoning_2117217c_v2,
  title = {document-level causal reasoning},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/document-level-causal-reasoning/rdr_2117217c43d342dcbaaad19f7325c4b6/v/2},
  note = {Version 2, frozen 2026-10-04. Accessed 2026-11-19},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake