曾元顯Tseng, Yuen-Hsien楊雪子Yang, Yukiko2025-12-092025-08-142025https://etds.lib.ntnu.edu.tw/thesis/detail/5fb108e14ca9724a18b4cd4c5ffa6193/http://rportal.lib.ntnu.edu.tw/handle/20.500.12235/124512本研究旨在探討如何運用多種人工智慧模型，對網路社群平台上的多標籤惡意文本（Toxic Content）進行分類與分析，並比較不同模型在多標籤分類任務中的成效。隨著社群媒體的普及，惡意評論與網路霸凌等問題日益嚴重，對使用者心理健康與社會互動造成負面影響。為有效偵測並管理有害內容，本研究選取來自Jigsaw釋出的Toxic Comment Classification涵蓋多重標籤的開放資料集，進行文本分析的實驗。研究將模型分為三大組：傳統機器學習模型組（如Logistic Regression、Random Forest、Naive Bayes、XGBoost等）、深度學習模型組（如GRU、BiLSTM、LSTM、CNN等）、以及大型語言模型組（BERT、Grok、GPT、Gemini）來進行分組實驗，訓練後模型的效能則依照 ROC-AUC、準確率（Accuracy）、F1-score、Hamming Loss 等指標來進行效能評估。實驗結果顯示，大型語言模型組的BERT在多標籤資料集的分類的表現最佳（ROC-AUC分數達0.9782），傳統機器學習中的 Logistic Regression搭配TF-IDF特徵次之，這可認為推出多年的傳統機器學習模型面對新推出的大型語言模型，效能表現仍相當亮眼，且無須額外費用，對學術或非商業的需求亦是理想的選擇，本研究結果可作為未來建立高效、精準之惡意評論自動分類系統的參考依據。This study aims to explore the application of various artificial intelligence models to the classification and analysis of multi-label toxic content on social media platforms, and to compare the performance of different models in multi-label classification tasks. With the widespread use of social media, issues such as toxic comments and cyberbullying have become increasingly severe, negatively affecting users’ mental health and social interactions. To effectively detect and manage harmful content, this study utilizes the publicly available Toxic Comment Classification dataset released by Jigsaw, which contains multiple labels to conduct text analysis experiments.The experiments in this study divide the AI models into three groups: traditional machine learning models (e.g., Logistic Regression, Random Forest, Naive Bayes, XGBoost), deep learning models (e.g., GRU, BiLSTM, LSTM, CNN), and large language models (BERT, Grok, GPT, Gemini) for comparative experiments. Model performance is evaluated using metrics including ROC-AUC, Accuracy, F1-score, and Hamming Loss. Experimental results show that BERT, from the large language model group, achieved the best performance on the multi-label dataset (ROC-AUC = 0.9782), followed by Logistic Regression with TF-IDF features in the traditional machine learning group. This suggests that even long-established traditional machine learning models can deliver competitive performance against newly developed large language models, without incurring additional costs, making them an ideal choice for academic or non-commercial purposes. The findings of this study provide valuable insights for the future development of efficient and accurate automated toxic comment classification systems.惡意評論機器學習深度學習大型語言模型Toxic commentsMachine LearningDeep LearningLarge Language Models學術論文