Most previous researches have focused on extracting event arguments from single sentences. However, in real-world scenarios, events are typically described across entire documents composed of multiple sentences. Thus, extracting event arguments at the document level is more practical. In this paper, we focus on Doc-level Event Argument Extraction (Doc-EAE), which presents two main challenges: (a) interference from irrelevant background information. Previous Doc-level approaches did not apply sentence filtering, treating all sentences as contributing equally to the event. This method overlooks the importance of sentences strongly associated with the event. (b) abstractness of role in template. In Doc-EAE, the mainstream approach relies on template-based argument generation. However, the roles defined in these templates often contain limited semantic information and are overly abstract, making them difficult to interpret. To address these challenges, we propose Key2Role, which contains two modules. The Template-based Key Sentence Selection (TKSS) module uses contextual attention weights between the template and document sentences to adaptively filter out sentences relevant to the event. The Role-guided Information Enhancement (RGIE) module leverages contextual role information from the document to build role ontology representations, providing potential role information to guide the selection of candidate spans. Results from the RAMS and WikiEvents datasets show that the proposed model achieved an increase of 0.7% F1 and 1.1% F1 improvement in overall accuracy.