Automated Software Requirements Prioritization using Natural Language Processing

Abstract

The software requirements specifications (SRS) may become a barrier to the successful completion of the project if they are written in a language that is difficult to understand. In certain situations, they cause failure to meet the actual requirements. The SRS dataset may contain redundant information or material that is disputed, either of which might result in higher expenditures and a loss of time, diminishing the overall efficiency of the project. The current developments in machine learning have led to a rise in the amount of work being put towards the development of automated solutions for the creation of a seamless software requirements specification (SRS). In this study, we employ the transformer models, including BERT and RoBERTa for classification. We focus on analyzing RoBERTa capacity for multiclass text classification tasks that involve predicting the type, priority, and severity of the requirements specified by the users. Moreover we compare its performance to that of other deep learning methods like LSTM and BiLSTM. We tested the performance of these models on the DOORS dataset. We have also compared the proposed model. We achieved higher accuracy i.e., ‘84.7%’, sensitivity, precision, recall and F1 score by using RoBERTa and compared our results with existing approaches.