Overcoming Data Shortage in Critical Domains With Data Augmentation for Natural Language Software Requirements

Abstract

Natural language processing (NLP) offers the potential to automate quality assurance of software requirement specifications. In particular, large-scale projects involving numerous suppliers can benefit from this improvement. However, due to privacy restrictions especially in highly restrictive industries, the availability of software requirements specification documents for training NLP tools is severely limited. Also, domain- and project-specific vocabulary, as such in the aerospace domain, require specialized models for processing effectively. To provide a sufficient amount of data to train such models, we studied algorithms for the augmentation of textual data. Four algorithms have been investigated by expanding a given set of requirements from the European Space projects generating correct and incorrect requirements. The initial study yielded data of poor quality due to the particularities of the domain-specific vocabulary, yet laid the foundation for the algorithms' improvement, which, eventually, resulted in an increased set of requirements, which is 20 times the size of the seed set. A complementing experiment demonstrated the usability of augmented requirements to support AI-based quality assurance of software requirements. Furthermore, a selected improvement of the augmentation algorithms demonstrated notable quality improvements by doubling the number of correctly augmented requirements.