BERT-Based Approach for Greening Software Requirements Engineering Through Non-Functional Requirements

The incorporation of sustainability principles during the requirements engineering phase of the development life cycle constitutes greening software requirements. This incorporation can have a variety of effects on the software design employed in modern and cutting-edge information technology (IT) systems. When sustainability principles are incorporated into requirements engineering, software design priorities can change and address current design issues such as energy and resource consumption, modularity, maintainability, and adaptability. In contrast to other green approaches that consider sustainable development, there is a further need to investigate the relationship between software development and the relevant green principles of sustainability during the requirements engineering phase. We present a new mechanism for mapping software nonfunctional requirements (NFRs) to defined dimensions of green software sustainability, consisting of two mapping steps: 1) between NFRs and sustainability dimensions; and 2) between sustainability dimensions and two clusters of green IT aspects defined in this work. The overall architecture of the promising approach is based on the use of the Bidirectional Encoder Representations from Transformers (BERT) language model with an expanded dataset. We consider transfer learning and domain-specific fine-tuning capabilities for constructing and evaluating a model specifically tailored for developing a proof of concept of the greening software requirements engineering task, as language models have recently emerged as a potent technique in the field of software engineering, with numerous applications in code analysis, automated documentation, and code generation. In addition, we test the model’s performance using an extended version of the PROMISE_exp dataset after adding a new binary classification column for categorizing sustainability dimensions into two defined clusters: Eco-technical and Socioeconomic, and having a selected domain expert label the raw data. The model’s efficiency is evaluated using four matrices—1) accuracy; 2) precision; 3) recall; and 4) F1 score—across a variety of epoch and batch sizes. Our numerical results demonstrate the viability of the approach in text classification tasks via performing well in mapping NFRs to software sustainability dimensions. This acts as a proof of concept for automating the sustainability measurement of software awareness at the early development stage. In addition, the results emphasize the importance of domain-specific fine-tuning and transfer learning for obtaining high performance in classification tasks in requirements engineering.