On the Use of GPT-4 for Creating Goal Models: An Exploratory Study

The emergence of large language models and conversational front-ends such as ChatGPT is revolutionizing many software engineering activities. The extent to which such technologies can help with requirements engineering activities, especially the ones surrounding modeling, however, remains to be seen. This paper reports on early experimental results on the potential use of GPT-4 in the latter context, with a focus on the development of goal-oriented models. We first explore GPT-4's current knowledge and mastering of a specific modeling language, namely the Goal-oriented Requirement Language (GRL). We then use four combinations of prompts – with and without a proposed textual syntax, and with and without contextual domain knowledge – to guide the creation of GRL models for two case studies. The first case study focuses on a well-documented topic in the goal modeling community (Kids Help Phone), whereas the second one explores a context for which, to our knowledge, no public goal models currently exist (Social Housing). We explore the interactive construction of a goal model through specific follow-up prompts aimed to fix model issues and expand on the model content. Our results suggest that GPT-4 preserves considerable knowledge on goal modeling, and although many elements generated by GPT-4 are generic, reflecting what is already in the prompt, or even incorrect, there is value in getting exposed to the generated concepts, many of which being non-obvious to stakeholders outside the domain. Furthermore, aggregating results from multiple runs yields a far better outcome than from any individual run.