Bridging Models and Language: An Encoder-Decoder Approach for Automated Architectural Documentation with LLMs

In modern software development, maintaining consistency between architectural documentation and implementation remains a significant challenge. This research explores how large language models (LLMs) can be integrated into an encoder-decoder framework to enable real-time architectural documentation and seamless conversion between semi-formal models (e.g., UML diagrams) and natural language descriptions. Building on previous work in LLM-based requirements verification, we develop a system that automatically converts models into text and vice versa. By allowing LLMs to assess textual descriptions for compliance with requirements, this approach reduces the need for extensive model-side validation through rule-based queries. Additionally, the decoder helps transform text-based process descriptions into structured architectural models, supporting digitalization in organizations. The system’s effectiveness is evaluated by comparing reconstructed models with their originals, assessing how well information is preserved and how accurately the transformations are performed.