FrameLLM for Requirements Generation: A Framework for Reducing Prompt Dependency and Improving Requirement Clarity and Completeness

Abstract

The adoption of Large Language Models (LLMs) in requirements generation offers significant opportunities for automatic requirements engineering. However, unstructured stakeholder inputs and heavy prompt reliance in LLM-based approaches can often result in inconsistent, incomplete, and ambiguous requirements. To address these challenges, the current study integrates NLP-based preprocessing with FrameNet semantics to guide LLM-based requirement generation. The evaluation considers both quantitative measures (number of generated requirements, execution time) and qualitative aspects (semantic completeness, domain relevance, consistency, and traceability). The initial results show that FrameNet-guided LLM (F_LLM) effectively generate clear, complete, and consistent requirements while reducing prompt dependency. Across multiple executions of five representative Autonomous Driving System (ADS) requirements, the proposed F_LLM consistently produced stable, semantically grounded, and traceable requirements, demonstrating improved completeness and efficiency compared to a baseline LLM. The final requirements are generated in the IEEE-830 standard format, making them ready for verification and implementation.