The Future is not One-dimensional: Complex Event Schema Induction by Graph Modeling for Event Prediction

Abstract

Event schemas encode knowledge of stereotypical structures of events and their connections. As events unfold, schemas are crucial to act as a scaffolding. Previous work on event schema induction focuses either on atomic events or linear temporal event sequences, ignoring the interplay between events via arguments and argument relations. We introduce a new concept of Temporal Complex Event Schema: a graph-based schema representation that encompasses events, arguments, temporal connections and argument relations.In addition, we propose a Temporal Event Graph Model that predicts event instances following the temporal complex event schema. To build and evaluate such schemas, we release a new schema learning corpus containing 6,399 documents accompanied with event graphs, and we have manually constructed gold-standard schemas. Intrinsic evaluations by schema matching and instance graph perplexity, prove the superior quality of our probabilistic graph schema library compared to linear representations. Extrinsic evaluation on schema-guided future event prediction further demonstrates the predictive power of our event graph model, significantly outperforming human schemas and baselines by more than 23.8% on

Schema-Guided Event Graph Completion

“…We conduct experiments on four event instance graph datasets: Car-Bombings, IED-Bombings, Suicide-IED, and Pandemic. The first three datasets are constructed by Li et al [15], which consist of complex events related to IED bombing. The last dataset Pandemic is constructed by us.…”

“…We use Disease-Outbreak schema for Pandemic dataset. The first two schemas are developed by Li et al [15], while the last schema is manually curated by us. The statistics of the four datasets and the three schemas are presented in Table 1.…”

“…The first class of event schema induction methods are sequence-based, which takes event-event relations into account, and orders event structures into sequences [8,28,29]. Since they fail to capture the multi-dimensional evolution of real-world complex events, i.e., an event can be preceded or followed by multiple events, researchers propose graph-based schema induction methods, which use graphs to formulate event schemas [15,16,27]. For example, Li et al [15] train an auto-regressive graph generation model on instance event graphs and then generate the event schema event by event.…”

“…However, note that a significant difference between event graphs and knowledge graphs is that, knowledge graphs are entity-centric and consist of static entity-entity relations, while event graphs are event-centric and dynamic, where new events and entities may appear as the graph evolves. Therefore, predicting missing events is a realistic, informative, and exciting task for event graph completion [15].…”

“…complex event template) is a generic and abstract representation of a specific type of complex events that encodes their stereotpyical structure. Event schemas can be either generated automatically by machines [8,15,16,28,29] from a large collection of historical event graphs, or manually curated by human. The lower part of Figure 1 illustrates the schema of general improvised explosive device (IED) bombing, where red and blue nodes represent the types of events and entities, respectively.…”