Abstract: To address the challenges of scarce annotated data and complex morphological structures in Tibetan event detection tasks, this paper proposes an end-to-end event detection method based on the CINO pre-trained model and a multi-level attention fusion mechanism. We construct a Tibetan government-comment event dataset of 2,012 samples with 10 event types. To effectively resolving Tibetan trigger word ambiguity and error propagation caused by word segmentation, we employ dynamic word segmentation boundary optimization strategies, fuse sentence-level and word-level semantic features and adopt a Conditional Random Field (CRF) for decoding constraints. Experiments demonstrate that our model significantly outperforms mainstream baseline models in Tibetan, achieving an F1-score of 85.07% in classification tasks and 77.79% in event detection tasks.