Generative AI for Business Process Management - Suitability of Modalities

Universität Ulm

MA Abschlussvortrag, Marvin Völter, Ort: O27/5202, Datum: 17.04.2024, Zeit: 10:00 Uhr

Business Process Management (BPM) is a thriving discipline that helps organizations run better. Previous work suggested using Large Language Models (LLMs) to significantly reduce effort in the discovery process of BPM by automatically generating process models from existing documentation. The documentation is often available in a multimodal form containing images and texts. This thesis presents an investigation of the capabilities of Generative Pre-trained Transformers (GPTs) to auto-generate graphical process models from multi-modal (i.e., text- and image-based) inputs. More precisely, it first introduces a small dataset for setting an objective common ground. Then, it applies multimodal GPT capabilities to generate process models using zero-, one- and few-shot prompting strategies. Moreover, it introduces an evaluation framework for the carried-out generation performance. The results indicate that GPTs can potentially be useful tools for semi-automated process modeling based on multi-modal inputs. More importantly, however, the dataset, evaluation metrics, and open-source evaluation code provide a structured framework for continued systematic evaluations moving forward.