Understanding event causality in natural language is critical for downstream applications such as decision making, policy planning, and risk assessment. While recent advances in Event Causality Extraction (ECE) have achieved promising results, most methods remain limited to extracting simple, sentence-level causal pairs, neglecting the complexity of real-world causal structures—such as multi-event chains, shared causes or effects, and cross-sentence relations. In this work, we propose a novel end-to-end framework that unifies event extraction and causal structure induction through a graph-based neural architecture. Our approach first encodes input documents with pretrained language models and heuristically initializes an event-centric relational graph based on domain-specific cues. We then introduce a relational Graph Neural Network (GNN) that dynamically updates event representations and relation types via multi-hop message passing. To address the challenge of extracting overlapping and implicit causal tuples, we design a decoding strategy that supports multi-causal relations and complex event interactions. Experimental results on a benchmark dataset demonstrate that our method outperforms state-of-the-art baselines by a significant margin in both precision and recall, particularly in multicause and cross-sentence scenarios. Our findings suggest that integrating structured relational reasoning into language models offers a robust path forward for comprehensive causal knowledge extraction.