Abstract: The task of few-shot event argument extraction requires the detection and classification of event-related argument entities with extremely limited training data, while traditional sequence labeling methods often struggle to transfer learned knowledge to new argument types in such scenarios. Upon analyzing this task, it was observed that argument entity detection involves detecting entities associated with given trigger words, and argument entity classification involves categorizing the relationships between trigger words and associated argument entities. Decoupling these tasks rationally can significantly reduce the strong dependence on event types.Based on this insight, this paper proposes a two-stage few-shot event argument extraction approach at the word level. Specifically, it introduces a combination of event type prompts, explicit trigger word prompts and implicit trigger word prompts to enable the model to detect trigger-related argument entities. When constructing a prototype network based on margin learning, the fusion of argument entity features and high-level conceptual features containing event type information simplifies the event type-related classification into a relationship classification based on trigger words and associated argument entities. Experiments on few-shot datasets constructed from either Chinese or English datasets demonstrate that this approach outperforms baseline models in both within-domain and cross-domain scenarios.