MNoR-BERT: multi-label classification of non-functional requirements using BERT

In the era of Internet access, software is easily available on digital distribution platforms such as app stores. The distribution of software on these platforms makes user feedback more accessible and can be used from requirements engineering to software maintenance context. However, such user reviews might contain technical information about the app that can be valuable for developers and software companies. Due to pervasive use of mobile apps, a large amount of data is created by users on daily basis. Manual identification and classification of such reviews are time-consuming and laborious tasks. Hence, automating this process is essential for assisting developers in managing these reviews efficiently. Prior studies have focused on classification of these reviews into bug reports, user experience, and feature requests. Nevertheless to date, a very few research papers have extracted Non-Functional Requirements (NFRs) present in these reviews. NFRs are considered as the set of quality attributes such as reliability, performance, security and usability of the software. Previous studies have utilized machine learning techniques to classify these reviews into their respective classes. However, it was observed that existing studies treat review classification problems as single-label classification problem, and also underestimate the contextual relationship between the words of review statements. To alleviate this limitation, the proposed research work used a transfer learning model to classify multi-label app reviews into four NFRs: Dependability, Performance, Supportability, and Usability. The proposed approach evaluates the performance of the pre-trained language model for multi-label review classification. In this paper, a set of experiments are conducted to compare the performance of the proposed model against the baseline machine learning with binary relevance and keyword based approach. We evaluated our approach over a dataset of 6000 user reviews of 24 iOS apps. Experimental results show that the proposed model outperforms state-of-the-art baseline techniques with respect to precision, recall, and F1-measure.