Efficient Extraction of Technical Requirements Applying Data Augmentation

Requirements for complex technical systems are documented in natural language sources. Manually extracting requirements from these documents – e.g., to transfer them to a requirements management tool – is time-consuming and error-prone. Today, machine learning approaches are used to classify natural language requirements and thus enable extraction of these requirements. However, in practice there is often not enough labeled domain-specific data available to train such models. For this reason, this work investigates the performance in artificially generating requirements through data augmentation. First, success criteria for a method for extracting and augmenting requirements are elicited in cooperation with industry experts. Second, the performance in the augmentation of requirements data is investigated. The results show that GPT-J is suitable for generating artificial requirements: weighted average F1-score: 62.74 %. Third, a method is developed to extract requirements from specifications, augment requirements data, and then classify the requirements. As a final step, the method is evaluated with requirements data from three industry case examples of the engineering service provider EDAG Engineering GmbH: assembly latch hood, adjustable stopper hood and trunk curtain roller blind. Evaluation shows that especially the transferability of models is improved when they are trained with augmented data. The developed method facilitates eliciting complete requirements sets. Performance of artificial intelligence models in requirements extraction is improved applying augmented data and therefore the method leads to efficient product development.