Detecting Non-Atomic Requirements in Software Requirements Specifications Using Classification Methods

Abstract

Requirements engineering is the most important stage in software engineering, one of which is to carry out specifications on requirements. Errors that occur at this stage will have a very bad impact on the next stages. A mistake that often occurs is a misunderstanding between stakeholders regarding the document specifications, and this is due to different backgrounds or fields of science. In addition, errors can also occur when making specification documents, for example, there are still non-atomic requirements in the document. Non-atomic requirements are a statement of requirements in which there is not only one element/function of the system. This research was conducted to develop a model that can detect non-atomic requirements in the software specification requirements written in natural languages. The initial stage of this research was to make a list of expert annotations (corpus) containing statements of atomic and non-atomic requirements. This Corpus later used as training data and test data in this study. Based on the corpus created, feature extraction and keyword generation carried out. The best model built in this research was produced by the classification method that used the Bayes Net algorithm. The result of the classification model was evaluated against human annotator using Cohen Kappa. The reliability of the model is considered fair for non-balance data in detecting non-atomic requirements in the software requirements specification. The reliability of the model is considered moderate for balance data in detecting non-atomic requirements.