Enhancing Requirements Engineering with Large Language Models: From Elicitation and Classification to Traceability, Ambiguity Management and API Recommendation

This thesis investigates the integration of Large Language Models (LLMs) into Requirements Engineering to enhance the overall management of software requirements. The research develops an extensive literature review that begins with the evolution of sequence modeling—from RNNs to the attention mechanism and encoder-decoder frameworks—and proceeds by examining a large number of the most recent approaches that apply Pre-Trained and Large Language Models to the key requirements engineering tasks—from elicitation and classification to ambiguity management, traceability and finally API recommendation. Along with this, a practical experiment on API tracing is conducted. The experiment employs a multi-step approach based on Goal-Oriented Requirements Engineering (GORE) to map low-level software goals—generated based on natural language project documentation—to specific API endpoints detailed in Swagger files. The experiment simulates the distinct tasks of LLM-based agents through iterative conversational interactions with GPT-4. The results highlight that while the model can effectively generate and map user-centric goals to API endpoints, performance is variable, with occasionally inconsistent mappings and omissions, which underscores the need for further interventions aimed at improving results, including enhanced prompt engineering, quality and clarity of software description and, ultimately, a structured approach leveraging an agentic architecture. Overall, this work contributes with an extensive overview of the recent LLM-based approaches to the core tasks of requirements engineering, while also offering practical insights through the experiment with the GPT-4 model in eliciting software goals and mapping them to API endpoints, showing challenges and benefits of applying LLMs to requirements engineering.