Applying machine learning algorithms to provide quality requirements specification

Abstract

This article is devoted to the problem of ensuring the quality of requirements specifications for complex technical systems. The purpose of this article is to use neural networks, classification and clustering algorithms to check requirements specifications for consistency and atomicity. It is believed that the use of neural networks will provide a vector representation of textual requirements formulations in order to identify inconsistencies in requirements specifications and to check the atomicity of individual requirements. This article demonstrates the use of such natural language processing techniques as fasttext, doc2vec, and BERT. K-means clustering is used to find inconsistencies in requirements specifications based on the assumption that the requirements of one cluster are potentially conflicting. Requirements are checked for atomicity by using gradient boosting over decision trees. The study showed that using the pretrained BERT neural network gives the best vector representations of requirements for solving clustering and classification problems using k-Means and gradient boosting, respectively. In addition, training the doc2vec model on requirements specifications is impractical, because the number of requirements in the specifications is usually limited and not enough for training, and FastText does not consider the semantics of the full requirement statement. In conclusion, a comparison of the results of the natural language processing methods considered in the article is given.