Identification of Architecturally Significant Non-Functional Requirement

Software requirements which are significant for designing Software architecture are called architecturally significant requirements (ASR). If ASR is not correctly identified, the resulting architecture will not be good. Wrongly designed software can't achieve the desired goal and quality, and this eventually lead to the complete failure of the software. Due to the complex behaviors behind architectural requirements, identifying the correct requirement is complex even for experienced architects. Identification and classification of ASR using machine learning algorithms have been reported in the past. However, their work didn't include Non-functional requirements (NFR) which have more impact than the ordinary NFR that have little effect on the architecture. The significancy of NFR vary from system to system. In this study, we have built a machine learning model for the identification of architecturally significant non-functional requirements (ASNFR) for a real-time system from the SRS document. The proposed model used three machine learning techniques: support vectored machine (SVM), Naive Bayes (NB), and K-Nearest Neighbor (KNN) using feature extraction techniques TF-IDF and software engineering pre-trained word2vec model. Grid search cross-validation techniques are used to tune the optimal value of hyperparameters of algorithms. We have prepared our own dataset and used 10 fold stratified cross-validation for evaluating and comparing the model. ASNFR identification model predicts 88% accuracy using SVM with TF-IDF and 87% in NB and KNN using TF-IDF and it predicts 73%, 70%, and 75% using SVM, NB, and KNN with pre-trained word2vec respectively. SVM with TF-IDF outperforms the others for the identification of ASNFR.