網路論壇中文諷刺意圖偵測

Abstract

本研究旨在尋找能有效幫助自動化諷刺偵測並具備語言學學理支持的諷刺特徵。本 研究收集以新冠肺炎為主題的網路論壇語料庫為對象，分析其中的留言內容，探討論壇 用戶如何洞察到諷刺言論的出現。若特定語言表達可以使論壇用戶察覺留言者諷刺他人， 則我們將此類語言表達稱為「諷刺特徵」。這些特徵可以分為留言層次諷刺特徵 (comment level) 和上下文層次諷刺特徵 (contextual level)。此外根據不同的提示諷刺手法，我們可以將不同諷刺特徵歸類。我們將這些諷刺特徵作為機器學習實驗分類重要依據。我們的結果顯示這些諷刺特徵對於建置諷刺檢測模型是有效的，並可提高詞袋模型 (bag-of-wordsmodel) 的表現。研究結果顯示，在辨識諷刺言論時，留言層次諷刺特徵和 上下文層次諷刺特徵同等重要。這表明我們的諷刺特徵具備兩種特性。第一，某些特徵 為特定語句，並與諷刺有著密切的關聯。第二，某些特徵則涵蓋前文內容，並藉由分析 前文語境來判斷發言者/作者的諷刺意圖。此外，我們也進一步發現了有效的諷刺特徵 及其策略。最後，在進行諷刺留言的情感分析時，我們注意到留言和其語境(如貼文或 先前留言)的正負情感對比有助於識別諷刺，並且發現諷刺不一定是在負面語境下的正 面話語，而是可能出現在正面語境下的負面/中立話語中。這一結果表明，在識別諷刺 時，人們應該要更注重不同的諷刺表現方式。本研究提供有效的諷刺特徵和提供諷刺標 註的程序和特徵選擇的依據。最後，本研究所提供的諷刺特徵和其策略歸類有助於辨識中文諷刺。This study aimed to investigate linguistically motivated cues for identifying sarcastic utterances automatically, using a COVID-related news corpus. The cues were categorized into comment and contextual levels. The linguistically motivated sarcasm cues could be also further categorized based on how they arouse the hearer's/reader's feelings that sarcasm was present. We applied these linguistically motivated sarcasm cues as input models in machine learning experiments to test their effectiveness. The results of the experiments showed that linguistically motivated sarcasm cues were effective in identifying sarcasm and that they are able to improve the performance of bag-of-words models. Results showed that both comment and contextual level cues were important for identifying sarcastic utterances, indicating the model's ability to identify utterances with a strong association with sarcasm and those that required prior context. Effective strategies and cues were identified. Moreover, polarity incongruity was found to be helpful in identifying sarcastic utterances, which were not necessarily positive in a negative context but could be negative or neutral in a positive context. This study contributes to identifying Mandarin sarcastic utterances by offering effective cues and outlining helpful procedures forsarcasm annotations and cue selection.