Automatic Classification of Requirements Based on Convolutional Neural Networks

The results of the requirements engineering process are predominantly documented in natural language requirements specifications. Besides the actual requirements, these documents contain additional content such as explanations, summaries, and figures. For the later use of requirements specifications, it is important to be able to differentiate between legally relevant requirements and other auxiliary content. Therefore, one of our industry partners demands the requirements engineers to manually label each content element of a requirements specification as "requirement" or "information". However, this manual labeling task is time-consuming and error-prone. In this paper, we present an approach to automatically classify content elements of a natural language requirements specification as "requirement" or "information". Our approach uses convolutional neural networks. In an initial evaluation on a real-world automotive requirements specification, our approach was able to detect requirements with a precision of 0.73 and a recall of 0.89. The approach increases the quality of requirements specifications in the sense that it discriminates important content for following activities (e.g., which parts of the specification do I need to test?).