The hierarchical structure of text, from articles to paragraphs to sentences, inherently reflects the compositional relationships between events at different levels, an important aspect often overlooked in prior research. In this work, we leverage this hierarchical structure to model the granularity of event types, thereby enabling more effective event feature construction for the task of event category discovery. We introduce a novel model, Multi-Dimensional Event Construction (MEC), which maps event type granularity to different layers of the encoder. The model extracts event features along two dimensions: layer depth to represent granularity and memory breadth to capture historical features. These features are then refined using Supervised Learning and Self-Contrastive Learning. Experimental results demonstrate that MEC achieves strong performance on two large-scale datasets, highlighting its ability to construct comprehensive event features and effectively discover unseen categories.