近年來，自然語言處理（Natural Language Processing， NLP）技術已廣泛應用於智慧客服、問答系統與資訊檢索。作為核心技術，句子語意配對（Sentence Semantic Matching， SSM）在將使用者輸入映射至既有意圖方面發揮著關鍵作用。然而，在特定領域情境中，SSM 仍面臨顯著挑戰：短文本中有限的語意線索導致語意稀疏、高度相似意圖間的細微差異導致表示邊界模糊，以及通用預訓練模型缺乏特定領域知識。 為解決這些挑戰，本研究提出 DAKA（Dual-Knowledge Augmented K-BERT with Angle Optimization）。具體而言，DAKA 引入雙重知識圖譜機制，在訓練期間利用外部關聯知識豐富語意上下文，同時在推論期間透過內部關鍵字知識提供決策約束，從而平衡語意擴充與預測穩定性。此外，為解決高度相似意圖間邊界模糊的問題，該模型將角度優化對比學習與 K-BERT 相結合，以有效分離表示空間中相似樣本的幾何邊界。最後，採用兩階段微調策略，以確保優化後的表示品質能平穩轉化為最終的分類決策。 為驗證 DAKA 的實用性，我們在真實客服資料（F-Transportation）和公開資料集（CLINC150）上進行了實證評估。實驗結果表明，DAKA 在所有評估指標上均持續優於各項基準模型。此外，個案研究進一步證實了所提機制在實際推論中的有效性，能夠精確修正高度模糊的樣本。整體而言，本研究證實知識的功能性拆分與協同應用能顯著提升短文本意圖辨識的準確率與推論可控制性，突顯了 DAKA 在特定領域客服情境中的實用價值。;In recent years， Natural Language Processing (NLP) techniques have been widely applied in intelligent customer service， question answering， and information retrieval. As a core technology， Sentence Semantic Matching (SSM) plays a crucial role in mapping user inputs to existing intents. However， in domain-specific scenarios， SSM still faces significant challenges: limited semantic cues in short texts leading to semantic sparsity， subtle differences among highly similar intents causing blurred representation boundaries， and a lack of domain-specific knowledge in general-purpose pre-trained models. To address these challenges， this study proposes a novel model， DAKA， which incorporates a dual knowledge graph mechanism with angle-optimized contrastive learning. We introduce a framework that leverages external and internal knowledge across training and inference to balance semantic expansion and prediction stability， effectively separating the boundaries of highly similar intents. Finally， a two-stage fine-tuning strategy is employed to translate the optimized representations into precise classification decisions. Furthermore， extensive experiments conducted on real-world and public datasets (F-Transportation and CLINC150) demonstrate the performance and superiority of DAKA over baseline methods across various evaluation metrics. We also use case studies to show the improvement and influence of the proposed mechanism on precisely correcting highly ambiguous samples during inference. Overall， this study verifies that the functional decomposition and synergistic application of knowledge significantly improve the accuracy and inference controllability of short-text intent recognition， highlighting DAKA′s practical value in domain-specific customer service scenarios.
