As our world grows in complexity, companies and employees alike need, more than ever before, solutions tailored to their needs. Such tools do not always exist and need to be designed from scratch. In this paper, we present a syntactic rule-based extraction tool for software requirements speci cation documents. Notably, our tool allows non-expert users to express their software needs in un ltered natural language.
With data sparsity being a major obstacle, modern approaches involving supervised or unsupervised machine learning methods are inconceivable. As a result, our architecture is based on a syntactic rule-based algorithm powered by a state-of-the-art dependency parser. In short, our approach aims at extracting every single subjectverb-object triple (SVO) from an input document. We consider this elementary approach to be an e cient and reliable way to extract actions and requirements from text and that, without the need for labelled data. Copyright c 2018 by the paper's authors. Copying permitted for private and academic purposes.
Given the extent of this research project, this study was divided into three phases. In this ongoing rst phase, the focus was set, as seen in Figure 1, on retrieving SVOs, relevant adjectives, complements, and negative words. Moreover, since the order of requirements found in a document may not always represent the true logical order, we experimented with di erent lexicon-based techniques as a means to sort all extracted software requirements into a proper sequential order. Lastly, to address the ambiguity dilemma, it was decided that supplementary semantic information about the individual requirements should also be generated using a word sense API.
In the second phase of this research project, the focus will be set on further re ning the syntactic rule-based extraction algorithm with the help of crowdsourced data collected during the rst phase. As it turns out, our tool will be made available to the public as a beta version; thus, allowing us to collect data and feedback. The acquired data will also be employed to build a classi cation model used for the validation of extracted software requirements. Since the SVO approach is purely syntactic, we are of the opinion that an authentication process should be added to the processing pipeline. Additionally, the possibility to replace the lexicon-based technique used to sort extracted requirements with a more contemporary solution will also be investigated.
Lastly, the third and nal phase will focus on solving common problems related to the eld of software requirements extraction, namely incompleteness, inconsistency, and vagueness [VK08]. As reported, requirements found in a speci cation document may be con icting or even lack some valuable information. Therefore, it is clear that this matter, too, needs to be studied. 3
Our approach aims at facilitating software development processes by giving (non-expert) users the opportunity to voice their needs and wishes in un ltered natural language. Because the tool's architecture is based on a syntactic rule-based approach, it has the advantage of not being domain-speci c, meaning that the origin of an input document should not a ect the overall performance of the extraction algorithm. As exposed, our research project will focus on overcoming common, yet arduous challenges related to software requirements engineering, namely ambiguity, incompleteness, inconsistency, and vagueness. Finally, we believe that the topic of sequential ordering is essential to any software requirements extraction task and will, therefore, be thoroughly examined.
This work was partially supported by the German Research Foundation (DFG) within the Collaborative Research Centre \On-The-Fly Computing" (SFB 901). [VK08]