為了解決上述問題，本研究提出穩健多標籤分類模型（Progressive Contrastive Learning with Self-Distillation， PCSD-MLC），整合三階段設計：首先，整合隱性增強機制與加權對比學習捕捉標籤間重疊的關聯；其次，導入分層學習率策略微調編碼器，減少訓練中底層語意被過度覆蓋；最後，整合多項優化策略，設計多代遞進式自蒸餾架構，使模型在不依賴外部大模型的前提下，持續自我萃取知識與提升效能。
為驗證PCSD-MLC的實用性，本研究以某大型企業真實搜尋資料以及公開法律文書資料集進行實驗。實驗結果顯示，PCSD-MLC於EM、Jaccard、Precision、Recall、Micro-F1與Macro-F1六項指標皆優於傳統序列模型、預訓練模型與多標籤專用架構等多種基準模型，並展現跨領域適應性；消融實驗也證實整合式對比學習、訓練優化策略與遞進式自蒸餾三者為效能的關鍵組件。個案研究以FDATA-2025所對應的企業真實搜尋場景為實證基礎，與企業現行LLM搭配RAG方案進行比較，結果顯示本模型在預測精準度、回應速度以及落地成本三個關鍵維度上全面領先，且敏感資料不外流，足以顯示模型的實務價值。;Driven by the trend toward service aggregation， super-apps have evolved to integrate diverse services—such as financial payments， retail， and telecommunications—into unified platforms. However， the rapid expansion of service nodes complicates user accessibility within these intricate environments. Consequently， this study focuses on developing an intelligent search mechanism tailored for such architectures to streamline service node connectivity and mitigate search bias.
Despite the integration of deep learning， intelligent search still suffers from semantic label overlap， where synonymous terms often correspond to distinct service intents， rendering traditional binary classification inadequate for fine-grained discrimination. Furthermore， the long-tail distribution of service nodes leaves sparse-sample tail labels challenging for models to effectively identify.
To address these challenges， we propose a robust multi-label classification framework termed Progressive Contrastive Learning with Self-Distillation (PCSD-MLC)， featuring a three-stage design. First， it combines an implicit augmentation mechanism with weighted contrastive learning to capture overlapping correlations among labels. Second， a layer-wise learning rate strategy is introduced to fine-tune the encoder， thereby preserving low-level semantics and preventing catastrophic overwriting during training. Finally， by integrating multiple optimization strategies， a multi-generation progressive self-distillation architecture is engineered to enable continuous knowledge extraction and performance enhancement autonomously， eliminating reliance on external large language models (LLMs).
To validate the efficacy of PCSD-MLC， extensive experiments were conducted on an empirical search dataset from a large enterprise alongside a public legal corpus. The experimental results demonstrate that PCSD-MLC consistently outperforms various baselines—including traditional sequential models， pre-trained language models， and specialized multi-label architectures—across six evaluation metrics: EM， Jaccard， precision， recall， Micro-F1， and Macro-F1， while exhibiting robust cross-domain adaptability. Ablation studies verify that integrated contrastive learning， training optimization strategies， and progressive self-distillation serve as pivotal components driving performance. Furthermore， a case study based on a real-world enterprise search scenario corresponding to FDATA-2025 compares the proposed framework against the enterprise′s incumbent LLM-RAG solution. The results indicate that PCSD-MLC achieves comprehensive superiority across three critical dimensions—prediction accuracy， latency， and deployment costs—while completely preventing sensitive data leakage， thereby establishing its substantial practical utility for enterprise application.
