When humans read a narrative, we naturally construct a multi-dimensional mental model of the events described. If we read that "heavy rains caused a flash flood, which subsequently destroyed a bridge," we instantly grasp three distinct dimensions: the temporal sequence (rain $\rightarrow$ flood $\rightarrow$ destruction), the causal links (rain caused flood; flood caused destruction), and the hierarchical structure (the bridge destruction is a sub-component of the broader flood disaster).
Historically, natural language processing and machine learning systems have treated these dimensions as isolated problems. As noted by Steven Bethard et al., traditional systems decouple temporal ordering and causal extraction into independent tasks. This fragmentation, also highlighted in a broad survey by Artuur Leeuwenberg and Marie-Francine Moens, creates a fundamental blind spot: it ignores the strict physical and logical laws that bind time and causality together.
The research direction of temporal-causal integration fixes this by modeling these dimensions jointly. By treating time, causality, and hierarchy not as separate tasks, but as interdependent facets of a single global structure, models can use the laws of physics as an inductive bias to resolve ambiguity and prevent logical contradictions.
The Intuition of Mutual Disambiguation
Why does joint modeling work so well? The core intuition relies on mutual disambiguation via physical constraints. The most unyielding constraint in our universe is that a cause must temporally precede or coincide with its effect.
Imagine an isolated causal extraction model trying to determine if Event A caused Event B. If the text is ambiguous, the causal model might guess incorrectly. However, if a temporal model is highly confident that Event B happened before Event A, an integrated system can immediately rule out the causal link. Conversely, identifying a causal relationship serves as a structural signal that instantly resolves temporal order.
[Isolated Models]
Temporal Model: Event A happened BEFORE Event B? (Uncertain)
Causal Model: Event B CAUSED Event A? (Uncertain)
[Joint Model with Physical Constraints]
Causal Link (B -> A) REQUIRES Temporal Order (B before A)
If Temporal Evidence shows (A before B), then Causal Link (B -> A) is IMPOSSIBLE.
Paramita Mirza formalized this in Extracting Temporal and Causal Relations between Events, demonstrating that enforcing this physical constraint prevents models from generating logically impossible timelines. To operationalize this mathematically, Qiang Ning et al. introduced a joint inference framework using Integer Linear Programming (ILP) in Joint Reasoning for Temporal and Causal Relations. By treating temporal and causal classifications as a globally optimized system constrained by physics, the accuracy of both tasks improves simultaneously.
This synergy is not limited to text. In computer vision, the authors of CaFlow: Enhancing Long-Term Action Quality Assessment with Causal Counterfactual Flow demonstrated that combining causal counterfactual reasoning with bidirectional temporal flow yields performance gains far exceeding what either component can achieve individually, mitigating both confounding bias and temporal instability.
Expanding the Graph: Hierarchies and Topology
Real-world events are rarely simple linear chains; they form complex networks. To fully exploit joint reasoning, researchers have expanded these constraints to include hierarchical (parent-child) and topological relationships.
Consider a subevent relationship: if "the bridge collapsed" is a subevent of "the flash flood," then the collapse must occur within the temporal boundaries of the flood. Haoyu Wang et al. leveraged this interplay in Joint Constrained Learning for Event-Event Relation Extraction by mapping these logical rules directly into differentiable loss functions. This allows a model to learn from cross-task dependencies even when training on separate datasets that lack complete joint annotations.
When moving to document-level analysis or complex event sequences, these interactions scale up:
- Document Logic: Meiqi Chen et al. showed in CHEER: Centrality-aware High-order Event Reasoning Network for Document-level Event Causality Identification that document-level understanding requires integrating temporal sequencing and hierarchical structures alongside causality.
- Narrative Generation: Xingwei Tan et al. argue in Cascading Large Language Models for Salient Event Graph Generation that understanding hierarchical containment is often a prerequisite for correctly mapping out temporal and causal links.
- Predictive Dynamics: Chao Liang notes in What Would Happen Next? Predicting Consequences from An Event Causality Graph that predicting future consequences accurately requires moving beyond isolated causal graphs to integrate temporal sequences and sub-event relationships.
This multidimensional integration also extends to structural data. Marc E. Maier developed a constraint-based algorithm in Learning the Structure of Causal Models with Relational and Temporal Dependence to capture causal structures across multiple interacting entities over time. In a similar vein, Yan Liu et al. and Ruichu Cai et al. demonstrated that combining temporal event histories with topological propagation networks provides a far more holistic and accurate view of how events influence one another across complex, interconnected systems.
From Feature Blending to Rationales
The methodology for capturing these intertwined relations has evolved from early feature engineering to modern deep learning and large language models (LLMs).
Early joint approaches, such as those by Kritika Venkatachalam et al. in SERC: Syntactic and Semantic Sequence based Event Relation Classification, focused on blending syntactic and semantic sequences from separate models to boost the performance of challenging causal classifications. However, these models were often bottlenecked by data scarcity. The creation of large-scale, unified benchmarks like MAVEN-ERE by Xiaozhi Wang et al. provided the million-scale data necessary to train deep neural models on coreference, temporal, causal, and subevent relations simultaneously, breaking the data silos that previously forced task isolation.
Today, the paradigm is shifting toward prompting LLMs, but the core requirement for joint logical consistency remains. Zhilei Hu et al. introduced Large Language Model-Based Event Relation Extraction with Rationales, which shifts the extraction process to a question-answering format that requires the model to generate explicit "rationales" (like transitive logic chains). By forcing the LLM to reason about temporal order and causal necessity simultaneously, the framework avoids the logical contradictions that plague independent pairwise extractions while reducing computational complexity.
Ultimately, joint reasoning works because it aligns a model's inductive biases with the structure of reality. Time provides the arrow, hierarchy provides the scale, and causality provides the mechanism. By enforcing their mutual dependencies, we transform event extraction from a series of disconnected statistical guesses into a coherent, logically constrained reconstruction of a narrative.
Key Foundational Work
Learning Semantic Links from a Corpus of Parallel Temporal and Causal Relations (2008) — contributed 1 concept elements to this direction.
Learning Temporal Causal Graphs for Relational Time-Series Analysis (2010) — contributed 1 concept elements to this direction.