Incorporating AI in the Teaching of Requirements Tracing Within Software Engineering

During the Software Development Lifecycle (SDLC), the first stage entails the Requirement Engineering phase. In this phase, engineers gather, analyze, and specify the requirements for a software system. Requirements playa crucial role in the SDLC as they establish the foundation for the entire system by defining the expected behaviors of the software system to be built. The resulting specifications are captured in a Software Requirement Specification (SRS) document. As part of the validation process, requirement specifications are traced. Requirement tracing involves linking the requirement to the artifacts where the customer requested the high-level requirement. Teaching proper requirements tracing can be challenging in a traditional classroom setting. It is essential to educate future software engineers on the proper process of developing an SRS document and of tracing requirements back to the originating artifact, which is also challenging due to the complexity and large scope of applying the complete requirements engineering process. Understanding how changes in customer needs can impact requirements is an imperative learning opportunity. In this work, we aim to incorporate the use of AI in the teaching of requirements tracing using Large Language Models. In this experiment, both GPT -3.5 and GPT -4 are provided the transcript of an interview between the customer and the engineering team, as well as the subsequent requirements elicited from that meeting and other customer provided artifacts. The GPTs are then instructed to determine which requirements can be traced back to the interview transcript. At the same time, the students (the requirements engineering team) conduct their own effort to trace requirements back to the original interview. The experiment was taken one step further to assess students' and the GPTs abilities to address requirements modifications. After another interview with the customer, where some needs were changed, some requirements were modified, and students, and GPTs were asked to trace the modified requirements to the new interview. The results proved that students are better than both GPT versions at tracing modified requirements, yet GPTs again identified requirements that students didn't trace back. The findings, illustrate that AI can help in the teaching of requirement tracing; these results suggest that while no AI model is currently capable of replacing real requirement engineers as they don't outperform students, it can be used as a tool to test the completeness of the requirement tracing process. We posit that GPT can be a tool for students to self-assess the degree to which their own requirements tracing is exhaustive.