Event logs obtained from modern information systems often record data at a very fine level of granularity. This can lead to traces containing hundreds of distinct low-level activities, resulting in process models that are far too complex for human comprehension. Empirical evidence suggests that analysts can effectively interpret models with at most a few dozen nodes, highlighting the need for higher-level abstractions. This challenge is even more pronounced in task mining, where detailed user click streams are analyzed. Bridging the gap between fine-grained events and meaningful process activities and/or tasks is therefore crucial for generating actionable insights. In this paper, we propose an iterative, hierarchical framework that combines pattern based activity-interaction graphs with community detection algorithms for clustering coherent sets of activities. These clusters are automatically labeled with concise, human-readable task names using LLMs, bridging the gap between structural abstraction and semantic interpretability. The resulting abstractions not only reduce model complexity but also enhance interpretability, enabling process mining to provide more meaningful and business-relevant insights.