CausalSense: Leveraging Common Sense Knowledge and LLMs for Joint Event Extraction and Relation Classification

Can a machine distinguish between what causes an event and what merely makes it possible?

CausalSense: Common Sense for Event Relations (LREC 2026)

Understanding events means more than detecting what happened. It also means recognising whether one event causes, enables, prevents, or intends another.

In this paper, we introduce CausalSense, a large-scale resource that combines news data, common-sense knowledge from ATOMIC, and synthetic examples generated with LLMs. The resulting dataset contains over 500,000 sentences annotated with fine-grained event relations.

Our models jointly identify event spans and classify their relations, improving over the previous state of the art across the full pipeline:

CausalSense shows how common-sense knowledge and data augmentation can help systems move beyond simple causality towards a richer understanding of how events are connected.