<p>Document-level event extraction is an important task in natural language processing, which faces the challenges of arguments-scattering and multi-events. The current mainstream method decomposes document-level event extraction into three subtasks: entity extraction, event type recognition, and argument-filling. However, existing methods flew consider the impact of writing style on event extraction, and have complex models with high training costs. To address these problems, this paper proposes a document-level event extraction model that incorporates both writing style and argument role relation. First, Roberta is used to encoding the document sentences to obtain semantic features, and BiLSTM-CRF is used to identify candidate entities. Second, the writing style is modeled as the relative distribution of argument roles in the document and is used to supervise the event type recognition results. Third, the argument-filling is simplified to the relation judgment between the argument roles. This method effectively addresses the problem of scattered arguments by constructing event clusters based on the relation of argument roles. It also deals with the multi-event problem through event cluster division. Compared with previous path expansion-based methods, experimental results show that the model has lower training costs and performs better in low training epochs.</p>