然而，企業內部客服資料具有標準問句類別多、標註樣本有限、部門資料量不均、問句表述多樣與跨領域語意干擾等挑戰，導致模型在少樣本與多部門情境下容易產生分類不穩定與相近類別誤判。為解決上述問題，本研究提出 Hi-CLKD（Hierarchical Contrastive Learning with Knowledge Distillation）架構，整合大型語言模型資料增強、問句與標準問句配對式對比學習、部門分類器、知識蒸餾、特徵對齊與協同預測機制，以提升句子語意配對分類效果與部署可行性。
本研究以企業真實內部客服問句建置 Hi-CSDS（Hierarchical Customer Service Dataset）作為主要驗證資料，並以 CLINC150 公開意圖分類資料集作為補充驗證。實驗結果顯示，Hi-CLKD 能提升整體標準問句分類效果，並透過階層式分流降低候選搜尋範圍，改善客服問句查找效率。消融實驗亦顯示，對比學習、知識蒸餾與特徵對齊皆對模型表現具有正向貢獻。整體而言，Hi-CLKD 可作為企業內部智能客服系統中兼具分類準確性、查找效率與部署可行性之解決方案。;With the development of global digital transformation and enterprise service digitalization， intelligent customer service systems have gradually expanded from external customer support to internal enterprise service scenarios， supporting employee needs in human resources， information technology， finance， and administrative processes. Recent advances in natural language processing and pre-trained language models have enabled systems to better understand the semantics of user queries， helping enterprises reduce manual search and repetitive responses while improving internal service efficiency.
However， enterprise internal customer service data presents several challenges， including a large number of standard query classes， limited labeled samples， imbalanced departmental data， diverse user expressions， and cross-domain semantic interference. These issues make classification unstable in low-resource and multi-department scenarios and increase the risk of misclassifying semantically similar classes. To address these challenges， this study proposes Hi-CLKD （Hierarchical Contrastive Learning with Knowledge Distillation）， a framework that integrates large language model-based data augmentation， paired contrastive learning between user queries and standard queries， a department classifier， knowledge distillation， feature alignment， and collaborative prediction to improve standard query classification performance and practical deployability.
This study constructs Hi-CSDS （Hierarchical Customer Service Dataset） from real enterprise internal customer service queries as the primary evaluation dataset and uses the CLINC150 public intent classification dataset as a supplementary evaluation dataset. Experimental results show that Hi-CLKD improves standard query classification performance in enterprise internal customer service scenarios and reduces the candidate search space and cross-domain interference through department-level routing. Ablation experiments further indicate that data augmentation， contrastive learning， knowledge distillation， and feature alignment all contribute positively to model performance.
