Stacked Ensemble Deep Learning for the Classification of Nonfunctional Requirements

Requirements engineering is the foundation for software quality. Defining the correct software requirements in the initial phases of the software development life cycle minimizes project costs and efforts. While functional requirements (FRs) define the software features, nonfunctional requirements (NFRs), such as availability, performance, security, and reliability are essential for the acceptance and deployment of the software. Understanding software requirements from different stakeholders is a tedious task. Manual investigation of the stakeholder needs may skip important NFRs. Thus, the need for automatic requirements classification techniques arose to eliminate the misinterpretation of stakeholder needs and to speed up the development process. Several machine learning approaches targeted the classification of NFRs. We explore the recurrent neural network, long short-term memory, and gated recurrent unit deep learning (DL) methods. We apply the random search technique for hyperparameter optimization. Further, we use stacked ensemble learning to enhance the classification by combining the strengths of the base models using support vector machine as a meta-learner. We use grid search to optimize the hyperparameters of the meta-learner. Further, we compare the stacked ensemble approach with the BERT language model. The proposed approach is evaluated on 914 NFRs gathered from two datasets. Our ensemble model achieved a weighted average precision, recall, and F1-Score of 0.91, 0.90, 0.90, respectively.