中文網路留言之幽默偵測

Abstract

本研究旨在探討利用深度學習模型(deep learning)對於台灣網路論壇—批踢踢實業坊(簡稱 PTT 論壇)上推文(論壇中對留言的稱呼)的中文幽默文本分類。文中結合了失諧理論(Incongruity Theory)、貶抑理論(Disparagement Theory)及釋放理論(Release Theory)並提出幽默是由違反行為(behavioral violation)及邏輯/溝通(logical/maxim violation)上的原則所形成。本研究於兩種層次的語境中找尋行為及溝通違規以進行幽默分類。第一種語境為推文裡的局部語境(local context)；第二種語境為整個文章及推文互動的全局語境(global context)。研究結果發現，相較於使用詞袋特徵(bag-of-words)的傳統機器學習模型，利用局部語境資訊的 BERT 模型可以提升模型表現。當 BERT 模型使用全局語境時，語境資訊的提取方式則對模型表現有不同影響。當模型提取原始的全局語境資訊時，模型表現沒有進步，而經過注意力機制對文章各部分進行重新賦權後，模型表現則有微幅提升。本研究亦從事後分析獲得幾項發現：一，就幽默推文來說，局部語境常出現某些討論主題。二，推文確實與文章某些部分有較緊密的連結。三、文章與幽默推文的連貫性較低，此發現支持了幽默裡的失諧現象，即透過轉換看待事物的視角來製造幽默。

This study explores deep-learning-based humor classification in Chinese online comments collected from the popular Taiwanese online discussion forum, PTT. It incorporates the incongruity theory, disparagement theory, and release theory to understand humor as a combination of behavioral and logical/maxim violations. The research focuses on two levels of context to capture the behavioral and maxim violations in humor classification: the local context within the comment itself and the global context related to the original post to which the comment responds. The findings indicate that the BERT model using local contextual information significantly improves the model performance compared to the traditional machine learning model using the bag-of-words features, while the approach to incorporate the global context has impact on the performance of the BERT model. While incorporating the original information from the global context has limited contribution to the task, reweighting global context by the attention mechanism has mildly improved the model performance. Post-hoc analyses highlight common topics emerging from local context in humorous comments, partial connection between posts and comments, and lower coherence between posts and humorous comments, supporting the concept of incongruity in humor and emphasizing the role of diverse perspectives.