CURE: Contextual Understanding and Reasoning for Document-level Event Causality Identification

Document-level Event Causality Identification (DECI) aims to determine whether a causalrelationship holds between events that may be in intra- or inter-sentences. Existing methodstypically rely on uniform graph-based representations, often overlooking that different eventsdemand distinct contextual cues. We propose CURE, a DECI framework that integrates comprehensivedocument-level understanding with event-specific reasoning. Specifically, CURE firstleverages large language models (LLMs) to derive unique contextual representations for eachevent from the same long document, ensuring that each event’s embedding focuses on its mostrelevant information. Next, it constructs a sparse event dependency graph that selectively linkseach event to a limited set of preceding events, inspired by human reading and forgetting patterns.CURE also adapts its inference strategy for intra- and inter-sentence event pairs. Experiments ontwo public datasets show that CURE outperforms state-of-the-art baselines.