基於具注意力機制類神經網路的反諷偵測

Abstract

自動反諷偵測，旨在使電腦理解人類反諷文字背後的真實意圖。過去的研究者 嘗試諸多人工抽取的複雜特徵，與各式經典機器學習方法。本研究探討如何通過詞 嵌入和深層神經網路將深度學習模型應用於此任務。本研究使用三種不同的深度 學習模型，分別為卷積神經網絡、循環神經網絡、和具注意力機制的循環神經網絡。 結果顯示具注意力機制的循環神經網絡在無脈絡的 Twitter 資料集和具脈絡的 Reddit 資料集達到最好的表現。此外，通過觀察由具注意力機制的循環神經網絡產 生的注意力權重向量，試圖窺見注意力機制如何幫助深度學習模型找出反諷語言 的文字線索。

Automatic Irony Detection refers to making the computer understand the real intentions of the human behind the ironic language. Much work has been done using classic machine learning techniques together with various features. In contrast to sophisticated feature engineering, this research investigates how deep learning can be applied to the Automatic Irony Detection task with the help of word embedding. Three different deep learning models, including Convolutional Neural Network (CNN), Recurrent Neural Network (RNN), and Attentive RNN, are explored. It shows that the Attentive RNN achieves the state-of-the-art performance on contextless and contextualized dataset. Furthermore, with a closer look at the attention vectors generated by Attentive RNN, an insight into how the attention mechanism helps find out the linguistic clues of ironic utterances is provided.