運用深度學習建立中文反諷辨識模型與驗證之研究

Research on Establishing and Validating Chinese Irony Recognition Models by Deep Learning - the Case of 2020 Taiwan Presidential Candidate’s Fan Page Corpus on FaceBook

由於網路上的酸民文化，許多網友在表達意見時，常會以「反諷」的方式表達意見。而隨著網路意見探勘、情感分析已逐漸普遍應用於許多領域的網路輿情分析，「反諷辨識」變得更加迫切，因為「反諷」表達會導致「情感分析」任務產生誤判。當前，中文反諷辨識的相關研究較少，且語料庫不足，導致中文反諷辨識任務難以順利進行。因此，本研究使用網路爬蟲蒐集2020總統大選候選人臉書粉絲專頁的網友留言評論，透過規則篩選與人工標記，取得1055筆反諷文本，建置中文反諷語料庫，並訓練、提出三個反諷辨識模型進行比較。經由實驗結果顯示，本研究提出的三個反諷辨識模型，不論在精準率、召回率、F1分數上皆有不錯的表現，整體辨識的準確率亦均可達到86%以上，有助於中文反諷辨識任務，並降低情感分析遇到反諷文本會預測失真的問題。

"Due to the troll-culture on the Internet, many netizens often express their opinions in an “irony” way. As Internet opinion mining and sentiment analysis have gradually been widely used in Internet for many fields, “irony recognition” has become more urgent, because the expression of “irony” will lead to misjudgments in the sentiment analysis. At present, there are few relevant researches on Chinese irony recognition, and the corpus is insufficient, which makes it difficult to carry out the Chinese irony recognition task effectively. Therefore, this research collected comments from the fan page of the 2020 presidential election candidates by web crawler. Through filtering by rules and manual marking, 1,055 ironic texts are obtained, a Chinese irony corpus is built, and three irony recognitions models are trained and proposed for comparison. The experimental results show that these three models have good performances in precision, recall, and F1 ratio. The overall recognition accuracy of each models can reach more than 86%, which is helpful for Chinese irony recognition and reduces the problem of distortion of irony texts in sentiment analysis."