In highly complex and flexible environments, event logs tend to exhibit high levels of heterogeneity, and clustering-based methods are candidate techniques for simplifying the mined process models from the process observations. To compensate for the information loss occurring during clustering, semantic information from event logs may be extracted and organized in the form of knowledge structures such as process ontologies using methods of ontology learning. In this article, we propose an overall computational framework for event log pre-processing, and then focus on a specific component of the framework, namely event log aggregation. We develop a detailed system architecture for this component, along with an implemented and evaluated research prototype SemAgg. We use phrase-based semantic similarity between normalized event names to aggregate event logs in a hierarchical form. We discuss the practical implications of this work for learning lower level process ontology classes as well as performing further process mining and analytics.