Software Requirements Classification Using Machine Learning Algorithms.

Area of Science:

• Software Engineering
• Natural Language Processing
• Machine Learning

Background:

• Accurate software requirements classification is crucial in software engineering.
• Existing methods require comparison of various feature extraction and machine learning techniques.

Purpose of the Study:

• To compare Bag of Words (BoW), TF-IDF, and Chi Squared (CHI2) for feature extraction.
• To evaluate Logistic Regression (LR), SVM, MNB, and kNN for requirements classification.
• To determine the optimal combination for classifying functional and non-functional requirements.

Main Methods:

• Utilized the PROMISE_exp dataset of labeled software requirements.
• Applied text normalization, BoW, TF-IDF, and CHI2 for feature extraction/selection.
• Implemented LR, SVM, MNB, and kNN for classification tasks.

Main Results:

• TF-IDF with LR yielded the best F-measure (0.91 for binary classification, 0.74 for NF classification, 0.78 for general classification).
• TF-IDF and LR combination outperformed other tested methods.
• SVM tied with LR in binary classification performance.

Conclusions:

• TF-IDF feature extraction combined with Logistic Regression is highly effective for software requirements classification.
• The study provides a reproducible benchmark for future research in requirements engineering.
• Further research will explore additional algorithms and precision improvement techniques.