| Summary: | The correct classification of requirements has become an essential task within software engineering. This study shows a comparison among the text feature extraction techniques, and machine learning algorithms to the problem of requirements engineer classification to answer the two major questions “Which works best (Bag of Words (BoW) vs. Term Frequency–Inverse Document Frequency (TF-IDF) vs. Chi Squared (<inline-formula><math display="inline"><semantics><msup><mi>CHI</mi><mn>2</mn></msup></semantics></math></inline-formula>)) for classifying Software Requirements into Functional Requirements (FR) and Non-Functional Requirements (NF), and the sub-classes of Non-Functional Requirements?” and “Which Machine Learning Algorithm provides the best performance for the requirements classification task?”. The data used to perform the research was the PROMISE_exp, a recently made dataset that expands the already known PROMISE repository, a repository that contains labeled software requirements. All the documents from the database were cleaned with a set of normalization steps and the two feature extractions, and feature selection techniques used were BoW, TF-IDF and <inline-formula><math display="inline"><semantics><msup><mi>CHI</mi><mn>2</mn></msup></semantics></math></inline-formula> respectively. The algorithms used for classification were Logist Regression (LR), Support Vector Machine (SVM), Multinomial Naive Bayes (MNB) and k-Nearest Neighbors (kNN). The novelty of our work is the data used to perform the experiment, the details of the steps used to reproduce the classification, and the comparison between BoW, TF-IDF and <inline-formula><math display="inline"><semantics><msup><mi>CHI</mi><mn>2</mn></msup></semantics></math></inline-formula> for this repository not having been covered by other studies. This work will serve as a reference for the software engineering community and will help other researchers to understand the requirement classification process. We noticed that the use of TF-IDF followed by the use of LR had a better classification result to differentiate requirements, with an F-measure of 0.91 in binary classification (tying with SVM in that case), 0.74 in NF classification and 0.78 in general classification. As future work we intend to compare more algorithms and new forms to improve the precision of our models.
|
|---|
