Using Recurrent Neural Networks for Classification of Natural Language-based Non-functional Requirements.

In software projects, non-functional software requirements (NFRs) are critical because they specify system quality and constraints. As NFRs are in natural language, accurately analyzing NFRs requires domain knowledge, expertise, and significant human efforts. Automated approaches that can help identify and classify NFRs can lead to reduced ambiguity and misunderstanding among software engineers, decreasing developmental costs and increasing software quality. This paper investigates the effectiveness of leveraging machine learning techniques to automatically classify various types of NFRs. Specifically, we develop and train a recurrent neural network model, which has been evaluated to be effective in handling sequential natural language text, to classify natural language NFRs into five different categories: maintainability, operability, performance, security, and usability. We evaluate and detail insights from the experimental study performed on two data sets that contain almost 1,000 NFRs. The experimental results show that this approach can classify NFRs with a precision average of 84%, recall of 85%, F1-score near 84%, and classification accuracy of 88% on the testing data set. As indicated by the results, applying appropriate machine learning techniques can help reduce manual efforts, eliminate human mistakes, facilitate the software requirements analysis process, and lessen developmental costs.