Defects such as the duality and the incompleteness in natural language software requirements specification have a significant impact on the success of software projects. By now, many approaches have been proposed to assist requirements analysts to identify these defects. Different from these approaches, this paper focuses on the requirements incompleteness implied by the conditional statements, and proposes a sentence embedding- and antonym-based approach for detecting the requirements incompleteness. The basic idea is that when one condition is stated, its opposite condition should also be there. Otherwise, the requirements specification is incomplete. Based on the state-of-the-art machine learning and natural language processing techniques, the proposed approach first extracts the conditional sentences from the requirements specification, and elicits the conditional statements which contain one or more conditional expressions. Then, the conditional statements are clustered using the sentence embedding technique. The conditional statements in each cluster are further analyzed to detect the potential incompleteness by using negative particles and antonyms. A benchmark dataset from an aerospace requirements specification has been used to validate the proposed approach. The experimental results have shown that the recall of the proposed approach reaches 68.75%, and the F1-measure (F1) 52.38%.
CATE: CAusality Tree Extractor from Natural Language Requirements
