Conflict Detection of Functional Requirements Based on Clustering and Rule-Based System With Multi-Agent Escalation
Conflicts in functional requirements are a common problem in software development. Conflicts often lead to inconsistencies and failures in the design or implementation stages. Although the combination of clustering and a Rule-Based System (RBS) can narrow the scope of conflict checking, its accuracy is not optimal when functional requirement sentences do not fall within the checking rules of the Rule-Based System. This study attempts to overcome this limitation with a multi-agent framework as an additional checking layer after Clustering and the Rule-Based System. Each agent has its own specific task, with the agents used being Lexical, Contextual, Logic, Natural Language Inference (NLI), Consensus, and Explainability. Experiments on two public datasets show that the proposed framework consistently outperforms the clustering and RBS baseline, achieving F1-scores of 0.828 on PURE (+5.3%) and 0.375 on OPENCOSS (+207% relative improvement), with false positives reduced by 81.2% on OPENCOSS, demonstrating the effectiveness of the multi-agent layer in handling cases that RBS alone cannot resolve.