Minimal Data, Maximal Impact: Language Model-based Pipelines for the Automatic Generation of Use Case Diagrams from Requirements
DS 133: Proceedings of the 35th Symposium Design for X (DFX2024)
Year: 2024
Editor: Dieter Krause; Kristin Paetzold-Byhain; Sandro Wartzack
Author: Simon Schleifer; Adriana Lungu; Benjamin Kruse; Sebastiaan van Putten; Stefan Goetz; Sandro Wartzack
Series: DfX
Institution: Engineering Design (KTmfk), Friedrich-Alexander-Universitat Erlangen-Nurnberg; AUDI AG
Page(s): 240-249
DOI number: 10.35199/dfx2024.25
Abstract
Striving for unique selling points leads to an increase in product requirements, which are prevalently written in natural language. In order to mitigate inherent ambiguities in these requirements specifications, methods of Model-Based Systems Engineering, like use case diagrams, can be utilized. However, creating model-based requirements is time-consuming. Thus, two novel pipelines for the automatic generation of use case diagrams are proposed and discussed with focus on reducing the amount of needed annotated training data. The first pipeline combines named entity recognition with active learning. The second pipeline utilizes generative large language models and prompt engineering. Both pipelines are exemplarily applied to a requirements specification from the automotive industry.
Keywords: Use case diagrams, Natural language processing (NLP), Large language models (LLM), Model-based systems engineering (MBSE), Requirements engineering (RE)