Requirements Acquisition through Dialogue with ChatGPT as SE

Abstract

Insufficient requirements acquisition due to insufficient capabilities of system engineers (SEs) is a serious problem in software development. Insufficient requirements acquisition from customers causes rework in downstream processes and delays in schedules. On the other hand, the emergence of generative AI, which can learn a wealth of knowledge from the Web and interact with customers based on that knowledge, may solve this problem. In this study, ChatGPT was used as an SE and the experimenter as a customer, and prompts were given to ChatGPT for requirements acquisition. The experimental results showed that ChatGPT was able to acquire about 42.2% of the total requirements. We also confirmed that ChatGPT's ability to acquire requirements decreases as the questions become longer, the number of non-functional requrement questions such as security increases, and ChatGPT intensively asks questions related to the keywords given in the initial requrements.