Document-Level Event Extraction (DEE) requires the extraction of corresponding events and information from documents containing multiple sentences. Compared to Sentence-Level Event Extraction (SEE), DEE faces key challenges such as loose entity associations across sentences, scattered event elements, and the dynamic structure of multiple events. This paper proposes a Dynamic Proxy Node and Structure-Aware Adaptive Matching model (DP-SAAM) for DEE, aimed at improving performance in complex scenarios through dynamic semantic modeling and matching strategy optimization. First, a dynamic proxy node generation mechanism is introduced to aggregate global document-level semantics via sentence-level attention weighting, adaptively generating proxy nodes to enhance contextual representations of diverse event types. Second, a Heterogeneous Graph Neural Network (HGNN) is constructed to model multi-granularity interactions among proxy, entity, and sentence nodes, capturing cross-sentence higher-order dependencies through relation-aware message passing. Finally, a Structure-Aware Adaptive Matching (SAAM) strategy is proposed, which dynamically selects between the greedy algorithm and the Hungarian algorithm based on event complexity and frequency, balancing efficiency and accuracy. Extensive experiments on the ChFinANN and DuEE-Fin datasets demonstrate significant improvements, where DP-SAAM achieves F1 scores of 83.4% and 76.3% respectively, outperforming state-of-the-art baselines (ProCNet) by 0.4–0.7% points. Ablation studies validate the essential contributions of all proposed modules, confirming their collective effectiveness in improving model performance. This approach enhances the extraction of structured event information, particularly improving event-based search and knowledge retrieval applications through its document-level semantic integration capability.