本研究以政治社群文本為研究對象，針對其同時涵蓋多個議題（Topic）與多個指涉 對象（Target）之語意特性，將文本分類問題建構為多輸出（multi-output）且多標籤（multi label）分類任務，並提出一套整合不同文字特徵表示方法與分類模型之分析架構，以系 統性探討各類方法於政治文本分類情境下之適用性與效能表現。在資料處理與標籤建構 方面，本研究以PTT政黑版作為語料來源，透過資料前處理與人工標註建立多標籤資料 集，並結合主題模型（LDA）進行資料導向之主題探索，以輔助Topic標籤設計；Target 標籤則採類型化方式建構，以提升標籤穩定性與模型泛化能力。在模型建構方面，比較 TF-IDF、Sentence Transformer 及 CKIP-BERT 三種文字特徵表示方法，搭配 Logistic Regression、SVM、XGBoost、MLP 及 Transformer fine-tuning 等分類模型，建構單一任 務（ Topic-only、Target-only）與多輸出聯合學習（Joint multi-output）實驗架構，並以 Micro F1、Macro-F1 及 Samples-F1 等多標籤分類指標評估模型效能。實驗結果顯示，CKIP BERT 於多標籤政治文本分類任務中整體表現優於傳統詞頻方法，尤其在語意隱含及上 下文依賴較強之文本情境下具有較佳分類能力；Target任務整體表現亦略優於Topic任 務，顯示指涉對象通常具有較明確之語意線索。至於Joint multi-output架構，其整體效 能與Single-task 相近，雖未帶來穩定且顯著之效能提升，仍能在不大幅犧牲分類效能下 同時完成多層次語意分類，展現其應用潛力。本研究提出結合多輸出與多標籤學習之政 治文本分類架構，並透過系統性實驗比較不同語意表示方法與分類模型之效能差異，補 足既有研究多聚焦於單一任務之不足，為政治輿情分析與多層次語意建模提供具體方法 與實證基礎。 關鍵字：政治文本分類、多標籤學習、多輸出學習、CKIP-BERT、PTT;This study focuses on political social media texts and addresses their semantic characteristics in which multiple discussion topics and referenced entities coexist within a single document. The text classification problem is formulated as a multi-output and multi-label learning task， and an analytical framework is proposed to compare different text representation methods and classification models for political text classification. Posts collected from the PTT HatePolitics board are used to construct a multi-label dataset through data preprocessing and manual annotation. Latent Dirichlet Allocation (LDA) is employed to assist Topic label design， while Target labels are constructed using a type-based categorization strategy to improve label stability and generalizability. Three text representation methods， namely TF-IDF， Sentence Transformer， and CKIP-BERT， are combined with Logistic Regression， Support Vector Machine (SVM)， XGBoost， Multi-Layer Perceptron (MLP)， and Transformer fine-tuning under both single-task (Topic-only and Target-only) and joint multi-output settings. Model performance is evaluated using Micro-F1， Macro-F1， and Samples-F1. The results show that CKIP-BERT outperforms traditional frequency-based methods， particularly for texts with implicit semantics and strong contextual dependency. The Target task also achieves slightly better performance than the Topic task， indicating clearer semantic cues. Although the joint multi-output framework does not consistently outperform single-task learning， it performs multi-level semantic classification with comparable effectiveness， demonstrating its potential for political text analysis. This study proposes a unified framework integrating multi-output and multi-label learning， providing an empirical basis for political opinion mining and multi-level semantic modeling. Keywords: Political Text Classification， Multi-label Learning， Multi-output Learning， CKIP-BERT， PTT
