Abstract: Current methods of selecting in-context examples mostly involve simply selecting top-k examples based on semantic similarity with the test input. They have two drawbacks: firstly, they may select almost identical examples, thus unable to provide additional information; Secondly, they ignores event attributes, resulting in the selected examples being unrelated to the type of test event. Therefore, this paper proposes the three elements of selecting in-context examples in event argument extraction tasks: semantic similarity, example diversity, and event relevance. Furthermore, the TRGQA (Tripartite Retrieval-augmented Generative Question Answering) framework is proposed, which first selects examples comprehensively from these three elements and extends the selected most similar question answering pairs as prompts to the context of the current test input. Then, a generative model is used to finally decode event arguments into answers. The experiments on the ACE2005 dataset have demonstrated the effectiveness and interpretability of this method.