Detecting contradictions in software system requirements using artificial intelligence

Abstract

The paper addresses the detection of semantic contradictions in textual software requirements. Such contradictions occur when separate statements prescribe incompatible rules for the same action under the same conditions. They can be difficult to identify when requirements use different wording, include exceptions, or rely on implicit business rules. The study aims to define an analysis procedure that helps a requirements analyst identify and examine potentially conflicting pairs. The procedure extracts the actor, action, object, modality, applicable conditions, and numerical constraints from each requirement and then selects pairs related to the same function. Selection also considers whether requirements belong to the same process and whether their versions are current, since similar statements may describe different system states. A large language model is used to compare their meanings, while explicit rules check numerical boundaries and negation. Particular attention is paid to cases where a permission and a prohibition prove incompatible only after their time conditions or threshold values are compared. Each proposed finding must include references to the original passages, the shared condition, and an explanation of the possible conflict. This presentation makes the analysis traceable and allows the analyst to determine whether relevant context was lost during comparison. A medical appointment cancellation example demonstrates that similar wording alone cannot establish a contradiction without considering the scope of each rule. When the scope or exceptions remain unclear, the pair should be marked for clarification rather than treated as a confirmed contradiction. The analyst remains responsible for the final decision on whether the requirements are consistent. Once a conflict has been confirmed, the analyst should record the agreed rule and identify the specification statements that require revision. For future empirical assessment, the paper proposes an expert-annotated set of requirement pairs and the measurement of precision, recall, false-positive rate, and time needed to review a flagged pair. No quantitative performance results are reported at this stage.