近年來，問句匹配已成為現代對話系統中的重要核心技術，然而實際問答系統常因細粒度意圖重疊、口語化表達與領域標註資料不足等因素，導致句子語意匹配效能受限。本研究提出 KCL-SSM，一套結合教師–學生式對比學習、領域知識圖譜增強候選檢索與混合式評分機制之句子語意匹配框架。透過知識蒸餾、資料驅動式圖譜建構與語意–結構融合策略，提升模型對複雜問句之辨識能力。本研究於多個公開資料集與實際產業資料集進行實驗驗證，結果顯示 KCL-SSM 相較基準模型皆能取得更佳匹配效能。學術上，本研究提供結合語意學習與結構化知識之問句匹配方法，拓展領域型語意匹配研究。管理上，本研究可協助企業降低知識維護成本，提升智慧客服系統之可靠性與應用效率。;Recently， question matching has become a fundamental technology in modern dialogue systems. However， practical question-answering systems often suffer from limited sentence semantic matching performance due to fine-grained intent overlap， colloquial expressions， and insufficient domain-specific labeled data. This study proposes KCL-SSM， a sentence semantic matching framework that integrates teacher–student contrastive learning， domain knowledge graph-enhanced candidate retrieval， and a hybrid scoring mechanism. Through knowledge distillation， data-driven knowledge graph construction， and semantic–structural fusion strategies， the proposed framework improves the ability to identify complex user queries. Experiments are conducted on multiple public datasets and a real-world industrial dataset， and the results demonstrate that KCL-SSM consistently outperforms baseline models in matching effectiveness. From an academic perspective， this study provides a sentence semantic matching approach that combines semantic learning and structured knowledge， extending research on domain-specific semantic matching. From a managerial perspective， this study helps enterprises reduce knowledge maintenance costs and improve the reliability and efficiency of intelligent customer service systems.
