本研究聚焦於人工智慧（AI）在智慧客服系統中的應用，旨在開發一套能自動分類客服工單的深度學習模型，藉此提升企業在客服作業流程中的效率與準確性。隨著數位轉型加速，以及自然語言處理（NLP）技術的快速發展，傳統仰賴人工進行工單分類的方式已難以滿足高效率、低成本與高準確度的實務需求，特別是在面對短文本、語意模糊與語言多樣性等挑戰時，更顯不足。 本研究採用 BERT、RoBERTa、BART 等預訓練語言模型作為基礎，導入 Fine-Tuning 機制進行模型微調，並以實際企業客服資料進行訓練與測試，建立「客服工單分類器（CSTC）」模型系列。實驗評估包含分類準確率、精確率、召回率與 F1-score 等指標，並進行消融實驗與不同模型間的比較，以驗證模型效能。各類深度語言模型皆能有效處理非結構化、短篇幅的客服敘述文本，展現優異的語境理解與分類能力。針對實際部署的可行性提出建議，包括分詞器策略、學習率敏感度、Batch Size、Dropout 設定等參數優化方式，提供企業導入智慧工單分類模組的實務參考。 本研究不僅建立了可行的智慧客服分類解決方案，更實證其在真實企業資料中的應用潛力，有助於企業降低人力成本、提升客戶滿意度與服務反應速度，對於推動客服系統智慧化與自動化具有高度實用價值與參考意義。 ;This study explores the application of artificial intelligence in smart customer service systems, with a particular focus on the automatic classification of customer support tickets. As digital transformation accelerates and natural language processing (NLP) technologies advance, traditional manual ticket classification methods can no longer meet the demands for efficiency and precision. In this research, we fine-tune pre-trained language models such as BERT, RoBERTa, and BART to build multiple classification models, which are evaluated using real-world customer service datasets. Key performance metrics include accuracy, precision, recall, and F1-score. The results demonstrate that these deep learning models effectively understand short texts and ambiguous semantics, significantly improving classification accuracy and operational efficiency. Furthermore, the study analyzes the adaptability and scalability of different models in handling linguistic variation and data imbalance. Practical recommendations are provided for deploying these models in real-world enterprise environments. Overall, this research contributes to enhancing the level of automation in customer service operations, reducing labor costs, and improving customer satisfaction and response speed. The findings serve as a practical reference for the implementation and development of intelligent customer service systems in industry contexts.
