Classification and evaluation of non-functional Turkish software project requirements using transfer learning methods

both project success and software quality. Non-functional requirements, in particular, play a significant role in determining quality attributes such as security, performance, and usability. However, the expression of these requirements in natural language and their diverse structures make automatic classification a challenging task. This thesis aims to classify non-functional requirements from Turkish software projects using transfer learning-based natural language processing techniques. For this purpose, a unique dataset consisting of 3,640 requirement statements was manually labeled and compiled from 40 different software projects. The dataset was split into 80% training and 20% test data, and experiments were conducted using machine learning algorithms, deep learning models, and transformer-based architectures. Among the tested algorithms, Linear SVC outperformed other machine learning methods, CNN performed best among deep learning models, and BERTurk yielded the highest performance among transformer models. Overall, BERTurk achieved the best classification performance with an accuracy rate of 93.44%, making it the most successful model in the study. This work contributes to the intersection of software requirements engineering and natural language processing and provides a meaningful baseline for future studies focusing on the classification of Turkish non-functional requirements.