This thesis investigates whether embedding-based semantic abstraction applied before process discovery can improve the interpretability and coherence of discovered BPMN models, and how this preprocessing step affects established discovery-quality metrics.

Following a Design Science Research approach, the thesis develops an embedding-based preprocessing pipeline that proposes and validates consolidations of potentially related activity labels before applying standard discovery techniques. The artifact is grounded in practitioner input from expert interviews and evaluated using a real-world manuscript-handling event log. The evaluation compares a baseline discovery path against a semantically enhanced path using the same discovery algorithm (Split Miner) and assesses the results using established process mining metrics, including fitness, precision, generalization, and simplicity. These quantitative results are complemented by domain expert feedback on grouping quality and information preservation.

The results show that the analyst-guided semantic abstraction produced structurally simpler BPMN models, with 18.2% fewer activities and 17.2% fewer nodes, and a more consistent activity vocabulary. Precision remained unchanged, while generalization increased slightly and replay fitness decreased. These findings indicate a trade-off between structural simplification and replay fitness, suggesting that embedding-based abstraction can support interpretability-oriented process analysis when candidate groups are manually reviewed. The thesis contributes an operationalized, semi-automated semantic abstraction pipeline and descriptive empirical evidence on its effects on process model quality.