HIGR: Hierarchical Iterative Graph Reasoner for Document-Level Event Causality Identification

Event Causality Identification aims to detect causal relations between events in text, which depend on both local cues and global discourse structure. Existing document-level approaches typically perform uniform reasoning over all event pairs, overlooking two key facts: (1) intra-sentence causal relations are easier to recognize due to explicit local markers, and (2) confidently identified causality can provide structural information that benefits subsequent predictions. We propose the Hierarchical Iterative Graph Reasoner (HIGR), which explicitly leverages these observations. Within each iteration, HIGR first resolves intra-sentence causality to establish reliable local causal contexts and then uses them to support inter-sentence reasoning. Across iterations, HIGR progressively refines event representations by aggregating the causal structures identified so far, enabling increasingly informed predictions. We also introduce adaptive aggregation mechanisms that regulate information flow at both the stage and edge levels. Experiments show HIGR outperforms state-of-the-art methods on two datasets, with particularly significant improvements on challenging inter-sentence relations.