Identification and Classification of Non-Functional Requirements from User Reviews

Abstract

Nowadays software applications are used for diverse purposes. With the explosion of the web and mobile experiences, system design fully depends on who is using the application. For the success of the application, developers need to be aware of the users’ concerns and expectations of the application. Existing research investigated that user reviews or feedback contain the quality concerns of the software that should be considered as non-functional requirements of the software to deliver a high-quality product. User reviews are usually short, unstructured, and written in an informal language, thus making it challenging to classify them based on the NFR standards and the huge quantity makes it tedious to identify requirements from review in the first place. To resolve this, we used Transformer-based language models to automate the detection of requirements from user reviews and classify Non-Functional requirements into seven sub-classes. We compared the classification results of the BERT and RoBERTa models using various evaluation metrics, and found that the fine-tuned version of RoBERTa surpassed BERT in classifying user reviews into requirement and non-requirement classes, as well as in classifying Non-functional requirements into seven subclasses.