Event Extraction (EE) is a critical task in information extraction, comprising two subtasks: Event Detection (ED) and Event Argument Extraction (EAE). Recently, generation-based models, which excel at semantically complex tasks, have led to the emergence of high-performing Sequence-to-Sequence (Seq2Seq) approaches for EE. However, these methods encounter challenges in both subtasks. In ED, frequent event types tend to dominate the model's attention, thereby diminishing its ability to detect infrequent long-tail events in the same sentence. In EAE, models are prone to misidentifying entities irrelevant to the target event, especially in complex textual contexts. To address these issues, we introduce a proposal-based reformulation that decomposes EE into proposal generation and selective classification. We propose Keyword Proposal-based Event Extraction (KPEE), a novel two-stage framework that first generates keyword proposals, and then learns to classify each through single-choice prompting. This design mitigates the attention bias caused by dominant event types in ED, improving the detection of long-tail events. In EAE, it also enables the model to better distinguish irrelevant entities through explicit learning on false argument proposals. Experiments on the ACE2005 and ERE datasets demonstrate the effectiveness of KPEE, with improvements of 2.5%/1.0% in ED and 1.1%/1.9% in EAE respectively, compared to state-of-the-art methods.