10.29327/1588952.28-11
The Softgoal Interdependency Graph (SIG) models non-functional requirements (NFRs) by representing softgoals and their interrelationships. However, building a SIG is challenging as it requires a deep understanding of qualitative concepts that vary across domains. The Transparency SIG (TSIG), which integrates over 30 related qualities, exemplifies this complexity. This study explores whether Large Language Models (LLMs), specifically ChatGPT-3.5 and ChatGPT-4o, can augment the knowledge of the TSIG. Through interactive dialogues, we analyzed the models’ ability to suggest relevant content and structure. Our findings show that, using the TSIG as the Gold Standard, the ChatGPT-generated models demonstrated the ability to approximate the expert knowledge represented in the TSIG, as evidenced by three authors achieving over 84% recall. Furthermore, since precision varied significantly—from 29.4% to 100%—this highlights differences in the amount of false positives. These elements require further qualitative evaluation to determine which of them may actually contribute to augmenting the knowledge on transparency, as modeled by the TSIG.
Keywords: Large Language Models (LLMs); Non-Functional Requirements (NFR); Softgoals Interdependency Graph (SIG); Transparency; ChatGPT
@inproceedings{wer202510,
author = {Portugal, R. L. Q. and Silva, L. F. and Sousa, H. S. P. and Leite, J. C. S. P.},
title = {Exploring LLMs on Supporting the Elicitation of Knowledge for NFR Catalogs: Insights from the Transparency Case},
booktitle = {Anais do Workshop em Engenharia de Requisitos - Proceedings of the 28th Workshop on Requirements Engineering (WER2025)},
year = {2025},
issn = {2675-0066},
isbn = {978-65-01-52831-1},
doi = {10.29327/1588952.28-11}
}