LLM-Driven MDA Pipeline for Generating UML Class Diagrams and Code

Abstract

The transformation of textual specifications into formal software models is a major challenge in software design automation. This study presents an integrated approach that combines the natural language interpretation capabilities of transformer models with explicit concept structuring using a domain specific language (DSL). This DSL, designed as a pivotal intermediate layer, ensures continuity between semantic extraction, UML modeling, and automatic Python code generation. By following the model-driven architecture (MDA) paradigm, the proposed pipeline follows a structured progression from text to model to code while maintaining a high level of traceability and controllability. The experimental evaluation conducted on a dedicated annotated corpus demonstrates the accuracy of the models on UML entities and highlights the importance of DSL for validation, editing, and exploitation of results. This approach paves the way for the development of intelligent modeling tools and structuring of automated transformation chains.