Multi-Label Classification for Requirement Smell Detection in Natural-Language Requirements

Abstract

Natural-language software requirements often contain quality defects such as ambiguity, vagueness, subjectivity, and nonverifiability. This paper presents a multi-label approach for detecting five requirement smell categories: Subjective, Ambiguous, Nonverifiable, Negative, and Vague. The method adapts SetFit with a SentenceTransformer encoder, a weighted binary-relevance logistic-regression head, hint-aware augmentation, and label-specific threshold tuning. Experiments on harmonized datasets show that the approach supports automated requirement quality analysis and interpretable smell detection.