LLM-based Analysis of System Requirements

This dissertation investigates the application of Large Language Models (LLMs) in Requirements Engineering with an intention to solve typical issues of ambiguity, inconsistency, and incompleteness in natural language system and software requirements. Current approaches such as formal specification languages, controlled natural languages, and conventional Natural Language Processing (NLP) techniques, suffer from scalability or require domain expertise. The proposed approach provides two tasks for requirements analysis based on LLMs. The first is a prompt-based Named Entity Recognition (NER) for finding domain specific terms (e.g., functions, systems), and second a boilerplate-based reformulation of requirements as machine-readable structures that follows a context-free grammar. This allows for the transformation of vague stakeholder input into certain, analyzable, and testable specifications. The study conducts extensive experiments with a number of LLMs such as Hermes, Qwen, Mistral and Llama, on Star Tracker dataset, in terms of entity extraction. While, also, evaluate the performance of GPT 4.1 in boilerplate generation quality on Orbit Control System and Star Tracker datasets. Results indicate that LLM ensembles with prompt engineering provide the same NER performance as the individual LLMs, while improving other labels on precision or recall, and GPT 4.1 can produce consistent structured requirements, as long as the original requirements aren’t in passive voice and are instructive enough. Lastly, the thesis showcase that LLMs can enhance requirements analysis quality and scalability to connect the natural language expressiveness with formal specification and to enable smooth integration with Model-Based Systems Engineering and automated development pipelines.