ReRule: Temporal Rule-Augmented Language Modeling for Causal Event Chain Completion

Complex web-based event sequences are essential for deciphering social dynamics and fostering positive social welfare outcomes computationally. Temporal knowledge graph completion and script event prediction have received attention for event sequence modeling. However, conventional knowledge graph models based on discrete quadruples fail to capture the global semantics and long-range dependencies of event chains, while script-based approaches often ignore the semantic roles and variations of entities within events. To address these limitations, we design a novel causal event chain completion task and propose a reasoning framework for key entity completion of query tuples in event chains. The framework first employs rule-guided filtering based on logical and temporal heuristics to prune the candidate space. Instruction-tuned generative models are then used to perform context-sensitive candidate generation and ranking. This hybrid design enables both factual consistency and flexible generalization. Our framework supports zero-shot reasoning as well as fine-tuned settings for domain-specific adaptation. We also construct three dedicated datasets for this task and conduct extensive evaluations. Experimental results demonstrate that our approach outperforms existing baselines across multiple metrics, highlighting its effectiveness in structured event modeling and entity inference.