隨著網路社群平台的發展，諷刺文本在線上溝通中扮演了重要角色，但由於其隱晦特性及特殊表達方式，自動檢測諷刺文本在自然語言處理領域仍是一項挑戰。本研究旨在探討適用於繁體中文的諷刺文本自動檢測方法，透過結合多種先進的預訓練語言模型並採用集成學習策略，以提升識別準確性。為了研究當前台灣網路環境中常見的諷刺表達方式，本研究從台灣網路論壇PTT的籃球版（BasketballTW）收集資料，開發了一個繁體中文的諷刺資料集。在資料集建構過程中，選擇合適的標註人員並評估標記一致性，以確保資料品質。而實驗結果表明，集成學習的策略在繁體中文諷刺文本偵測上能夠有效提升分類的效能，特別是結合多個預訓練語言模型的預測機率可以顯著提升模型效能，而結合語言模型的最後一層隱藏層嵌入向量的方法，以及將多個預訓練語言模型的預測機率結合手工設計特徵的方法，在效能提升上則相對有限。;With the development of online social platforms, sarcastic texts play an increasingly important role in online communication. However, due to their implicit nature and unique expression, automatically detecting sarcastic texts remains a challenge in the field of natural language processing. This study aims to explore methods for automatically detecting sarcastic texts in Traditional Chinese by combining various advanced pre-trained language models and adopting ensemble learning strategies to enhance detection accuracy. Data was collected from the basketball message board (BasketballTW), which is one of Taiwan′s largest online forum, PTT, to develop a dataset of sarcastic texts in Traditional Chinese. During the dataset construction, appropriate annotators were selected and the consistency of annotations was evaluated to ensure data quality. Experimental results indicate that ensemble learning strategies significantly improve the classification performance of detecting sarcastic texts in Traditional Chinese, especially when combining the prediction probabilities of multiple pre-trained language models. However, the method of combining the last hidden layer embeddings of language models and integrating manually designed features with the prediction probabilities of multiple pre-trained language models shows relatively limited improvements in performance.
