Towards Process Mining with Natural Language Event Logs

Process mining is a well-established discipline concerned with extracting knowledge from process event logs, enabling analysis and improvement of business processes. Common use cases include process discovery, where process flow models are generated; conformance checking, where actual processes are compared against prescribed models; and process enhancement, which involves identifying bottlenecks and inefficiencies to optimise workflows.

At the core of process mining are event logs, which are structured records of events associated with specific process instances. Typically, each event in a traditional event log is unambiguously labelled with the process activity it represents, making it straightforward to analyse the flow of tasks.

However, many modern systems, such as Customer Relationship Management (CRM) systems or Patient Information Systems (PIS), support business processes in highly flexible ways, documenting process-related events using natural language rather than structured labels. These natural language event logs are not explicitly labelled with the process activity they represent, yet this information can often be inferred from the text, implying the possibility of applying process mining techniques to natural language event logs, although it presents significant challenges.

Labelling natural language events with the correct process activity is not straightforward. Depending on the granularity of the desired process model, two events might be mapped to the same process activity or to different activities. This thesis introduces the concept of a taxonomy of activities, which groups similar activities hierarchically to facilitate the mapping of natural language events to process activities at varying levels of detail.

Building on this, a novel process discovery framework is developed to work on natural language events, enabling the creation of hierarchical process models. These models allow process analysts to explore the process at different levels of abstraction interactively, selecting specific process activities from the taxonomy.

Additionally, a realistic, artificially generated dataset of natural language events is created, labelled with ground truth at multiple levels of abstraction. Finally, a case study in the mental healthcare domain demonstrates the practical applicability of the hierarchical process miner to real-world data, showing how it can reveal insights into complex, flexible processes in real-world settings.