A Thesis Submitted to the School of Graduate Studies in Partial Fulfillment of the Requirement for the Degree of Master of Science in Software Engineering

Abstract

Requirement engineering is a fundamental process in software development, focusing on identifying, documenting, and managing the requirements of a software system. It serves as a critical bridge between the stakeholders' needs and the final software product, ensuring that the developed system meets the expectations and functional requirements of its users. It is an integral process to the success of any software development project, as poorly defined requirements can lead to project failures, cost overruns, and user dissatisfaction. User feedback from Play Store reviews plays a crucial role after software deployment, offering valuable insights into user experiences, satisfaction, and areas for improvement. It helps developers understand how the product is being used and identify features that need attention or enhancement. However, previous studies have primarily focused on manual analysis or traditional machine learning techniques to extract requirements from user feedback. These methods often struggle with scalability, data complexity, and the ability to handle unstructured and imbalanced feedback data effectively. Furthermore, they cannot capture nuanced contextual information present in user feedback, leading to incomplete or inaccurate requirement extraction. This research gap highlights the need for more advanced and automated approaches to enhance the extraction process. Our study aims to address this challenge by developing a deep learning-based model to automatically classify user feedback into predefined software requirement categories. The objective is to develop a model that extracts relevant software requirements from user feedback that enhances the efficiency and relevance of requirement extraction by leveraging advanced deep learning models. The study explored five different deep learning models BERT, RoBERTa, XLNet, BiLSTM, and LSTM-CNN hybrid model, applied to a dataset of 10,009 user feedback collected from seven applications from the Google Play Store. The study followed an experimental research design to identify key problems such as the complexity of textual data, the imbalance in feedback categories, and the need for effective preprocessing techniques. To overcome these challenges, the study employed various strategies, including data balancing, label encoding, and hyper parameter tuning on different models. After extensive experimentation, the results demonstrate that pre-trained transformer models, particularly XLNet with an accuracy of 99.57%, outperformed other models in accurately classifying user feedback into software requirement categories. The findings suggest that incorporating advanced deep learning techniques can significantly improve the automation of software requirement extraction, leading to more efficient software development processes.