Automated BPMN Model Generation from Textual Process Descriptions: A Multi-Stage LLM-Driven Approach

A research team including Ion Matei, Maksym Zhenirovskyy, Praveen Kumar Menaka Sekar, and Hon Yung Wong has developed a scalable pipeline that uses Large Language Models (LLMs) to automate business process modeling. Their multi-stage approach tackles the long-standing challenge of converting unstructured natural-language process descriptions into standardized BPMN (Business Process Model and Notation) diagrams. The system first creates its own training data by translating multilingual BPMN XML files into English, validating them using execution-oriented compliance checks in the SpiffWorkflow engine, and then iteratively repairing any issues through targeted LLM-guided corrections. This produces a consistent, high-quality ground-truth corpus from which accurate process descriptions can be generated.

In their empirical study of 750 public BPMN diagrams, the pipeline successfully generated 387 validated ground-truth models. The reconstruction phase achieved an average similarity score above 0.75 against the original diagrams, with approximately 50 reconstructions being near-perfect matches that differed only in minor naming variations. The researchers introduced a sophisticated multi-dimensional similarity framework that combines structural metrics, type-distribution alignment, and embedding-based semantic measures to evaluate the quality of generated models. This demonstrates that LLMs can generate structurally compliant and semantically meaningful BPMN diagrams at scale, moving beyond simple pattern matching to true understanding of process semantics.

The approach represents a significant advancement in automated business process modeling, particularly because it handles the heterogeneity of modeling conventions and multilingual sources that have traditionally made this task difficult. By automating both ground-truth construction and model reconstruction, the pipeline reduces dependency on manually curated datasets and expert knowledge. The use of execution-oriented validation through SpiffWorkflow ensures that generated BPMN models aren't just visually correct but are actually executable workflows, adding practical utility to the academic achievement.