Gerando Descrições de Processo de Negócio a partir de Modelos BPMN: uma abordagem com Grandes Modelos de Linguagem
Year: 2024
Type: dissertation
Abstract: Business process models are effective ways of describing and documenting processes within an organization. However, discovering such processes in a Business Process Man agement (BPM) project is a challenging and time-consuming task. Domain experts are not familiar with process model notation, such as the Business Process Model and No tation (BPMN), so verbalization techniques must be employed for process validation. Automated techniques using Natural Language Processing (NLP) have been proposed for process description generation, and they have been proven effective and reliable. Gener ative AI is a burgeoning field that has shown unprecedented advances in the development of Large Language Models (LLMs) that have surpassed most state-of-the-art approaches for general-purpose problem-solving. Most recent LLMs are capable of complex rea soning, even with long inputs. In this study, we conduct a comparative study between traditional NLP process description generation techniques and a new approach based on LLMs. We investigate whether current LLMs are capable of generating a process descrip tion from a BPMN model. Through a complementary experimental study, we investigate the effects of process model complexity and prompting techniques on the resulting tex tual complexity. We conclude that current LLMs aren’t ready yet to be used for process description generation. Our experimental study showed that LLMs confidently generate process descriptions of up to 25 process elements and 17.3 sequence flows, after which an AI safety mechanism is triggered and a refusal is returned. However, our compara tive study presents empirical evidence that the LLM’s actual capabilities fall short of its self-assessment. For process models below this threshold, the generated description still contained errors. Finally, we establish a new research direction for the process description generation problem. We propose a hybrid approach that combines the structural correct ness of NLP techniques with the reasoning capabilities of LLMs.
Author: Gomes, Fernando Corrêa
Language: English
Subfield: Management Information Systems
Domain: Social Sciences
Open Access status: green
OpenAlex ID: W7000684610