<p>Document-level event extraction (DEE) confronts dual challenges in resolving interwoven multi-event structures and mitigating semantic fragmentation of arguments across sentences. While existing graph-based methods model local entity-event interactions through dense node connections, their oversimplified relational encoding has been found to propagate noise and obscure critical dependencies. A novel hierarchical semantic graph model termed CESG (comprehensive entity semantic graph) is presented, formalizing document-level semantics through dual-layer relationships: syntactic dependencies (document-sentence-token hierarchy) and conceptual interdependencies (event-argument-mention associations). The model’s innovation lies in its entity-centric relational architecture that precisely encodes intra-mention semantics and inter-node dependencies through structured edge definitions. A three-phase methodology enables precise event extraction through (1) joint entity extraction and embedding, (2) context-aware graph construction, and (3) graph neural network-based event encoding and decoding. Evaluations on financial document benchmarks (ChFinAnn, DuEE-Fin) demonstrate CESG’s consistent outperformance of existing methods, achieving absolute F1 gains of 0.4–1.0 points with efficacy in cross-sentence argument linking. These results establish that structured semantic unification of syntactic and conceptual relationships enhances robustness against information sparsity and contextual noise, offering a scalable solution for complex document processing in real-world scenarios.</p>