Abstract
Traditional event prediction models often treat real-world scenarios as linear sequences, failing to capture the intricate dependencies between participants and overlapping timelines. This paper introduces the Temporal Event Graph Model (TEGM), an auto-regressive framework that induces complex event schemas as multi-dimensional graphs. By modeling temporal order, participant roles, and entity relations simultaneously, TEGM significantly outperforms both human-curated schemas and state-of-the-art linear baselines in predicting future events.
The Limitations of One-Dimensional Modeling
In fields like intelligence analysis or disaster response, understanding the typical structure of a scenario—known as "event schema induction"—is critical. However, traditional approaches have largely been restricted to one dimension, viewing scenarios either as isolated "atomic" events or as simple linear chains (e.g., Event A leads to Event B).
Real-world crises are rarely so simple. A complex event like a car bombing involves a web of interconnected sub-events: a buyer purchasing a vehicle, an engineer assembling a device, and a driver attacking a target. These events are linked not just by time, but by shared "arguments"—the same person who buys the car might be the one who drives it. To accurately predict the next step in such a scenario, a model must understand the interplay between temporal sequences and participant roles.
The Temporal Event Graph Model (TEGM)
The authors propose TEGM, a framework that treats schema induction as a graph generation problem. Instead of predicting a string of text, the model "grows" a graph step-by-step through a five-stage process:
- Event Generation: The model predicts the most likely next event type (e.g., "Arrest") based on the current graph state.
- Edge-Aware Graph Neural Network (GNN): It uses a specialized GNN to update the graph's internal representation. Crucially, it applies different weights to different types of edges—temporal, argument roles, and entity relations—to preserve rich semantic meaning.
- Coreferential Argument Generation: Using a "copy mechanism," the model determines if a participant in a new event is a brand-new entity or an existing one (e.g., identifying that the "Detainee" in an arrest is the same "Attacker" seen previously).
- Entity Relation Edge Generation: It predicts the relationships between entities, such as whether a person is located in a specific geographic area (GPE).
- Event Temporal Ordering Prediction: Finally, the model predicts the specific timing dependencies between the new event and all existing events in the graph.
Experimental Setup and Results
The researchers evaluated TEGM using two primary datasets based on the DARPA KAIROS ontology: a General Schema Learning Corpus (82 event types across 6,399 documents) and a specialized IED Schema Learning Corpus focused on high-stakes improvised explosive device scenarios.
Key Findings
- Superior Predictive Power: In predicting the final event in a sequence, TEGM outperformed human-curated schemas and state-of-the-art language model baselines by more than 23.8% on HITS@1.
- The "Web" Beats the "Line": Results showed that modeling participant arguments is as vital as modeling the events themselves. Tracking how a specific entity moves through a narrative provides the context necessary to disambiguate future possibilities.
- Probabilistic Robustness: Interestingly, the AI model outperformed human-designed schemas. While human schemas describe "ideal" sequences, TEGM learns the probabilistic "messiness" of real-world data, allowing it to handle the noise and variations typical of actual reporting.
Conclusion and Future Work
By shifting from linear chains to multi-dimensional graphs, TEGM provides a more realistic and effective way to model complex real-world scenarios. The model's ability to represent branching futures—assigning probabilities to multiple potential outcomes—marks a significant step forward in predictive AI.
Future work will address current limitations, such as the model's sensitivity to errors in underlying Information Extraction (IE) systems, which accounted for over 40% of prediction errors. Additionally, the authors aim to explore hierarchical induction, where the model can understand "schemas within schemas" for even greater structural depth.
Summary updated