Data-driven Elicitation and Optimization of Dependencies between

Requirement dependencies affect many activities in the software development life cycle such as design, implementation, testing, release planning and change management. They are the basis for various software development decisions. However, requirements dependencies extraction is not only error-prone but also a cognitively and computationally complex problem that consumes substantial efforts, since most of the requirements are documented in natural language. This paper proposes a novel approach to extracts requirements dependencies utilizing natural-language processing (NLP) and weakly supervised learning (WSL) in two stages. In the first stage, binary dependencies (basic dependencies:dependent/independent) are identified, which are further analyzed to detect the type of the dependency in the second stage. An initial evaluation of this approach on the PURE data set - European Rail Traffic Management System - was carried out using three machine learners (Random Forest, Support Vector Machine and Naive Bayes), which were then compared and tested. Results showed that all the three learners exhibited similar accuracy measures, while SVM needed additional parameter tuning. The machine learners' accuracy was further improved by applying weakly supervised learning to generate pseudo annotations for unlabelled data. Based on these results, agenda is to provide decision support under a dynamic use case scenario that includes (i) continuous updates and analysis of dependencies, (ii) identification of the general types of dependencies, and (iii) dependencies as a key driver of the decision support for the product releases.