近年來，多標籤文本分類因成功被應用於多個領域而受到廣泛關注。該任務主要目標是為輸入文件指派多個相關標籤。為了進一步提升多標籤文本分類效能，許多研究將對比學習技術應用於此任務，並在一般情境下取得良好表現。然而，當文件語意較為多樣，或不同標籤對文件語意的貢獻程度不均衡時，現有方法仍面臨若干挑戰與限制。為了解決上述問題，本研究提出MCDL，一種以medotype (中心點) 為基礎的多標籤對比學習框架。研究中，我們引入中心點強化概念，以降低離群值對表示學習的影響，並提升表示結果的穩定性。MCDL主要由標籤表示學習與文件表示學習兩個模組構成。標籤表示學習模組將標籤頻率資訊整合至動態圖卷積網路中，以捕捉標籤之間的相依關係，並緩解長尾分布所帶來的影響。文件表示學習模組則結合標籤貢獻度與標籤文件注意力機制，以擷取不同標籤所關注的文本語意。最後，實驗結果顯示，相較於現有方法，MCDL在多個真實世界資料集上皆能取得更優異的表現。此外，本研究透過案例分析，進一步證實MCDL在真實應用情境中的實務性。;Recently， Multi-label text classification (MLTC) has attracted tremendous attention due to the successful applications in various domains， which aims to assign multiple relevant labels to a given document. Most prior contrastive learning-based methods could reach qualified performance on general MLTC task. However， when document semantics are diverse or label contributions are imbalance， existing methods still suffer several challenges and limitations. To address these issues， in this study， a novel medotype-based contrastive learning framework (MCDL) is developed for MLTC. We introduce a medotype-enhancing concept to reduce the influence of outliers and improve stability of the representation. MCDL mainly consists of two modules for label and document representation learning. The label representation learning module integrates frequency information with dynamic graph convolutional networks to capture label dependencies and alleviate long-tailed distributions. The document representation learning module adopts label-document attention with contribution mechanisms to capture label-specific document semantics. Finally， experimental results indicate that MCDL could achieve superior performance comparing to state-of-the-art methods on several real-world datasets. We also give a real case study to show the practicability of proposed MCDL.
