隨著社交媒體的發展，網路上充斥著大量的訊息。為了能夠快速地分析這些大量訊息中的正面及負面情緒，情緒分析已成為自然語言處理中重要的議題。在不同類型的情感分析中，諷刺偵測尤其扮演著極重要的角色。因為當一個句子含有諷刺意味時，其表面含義將會與其核心表達意思相反。為了避免錯誤的判斷，本研究提出了兩種提高諷刺偵測準確度的技術：構建輔助句與堆疊多個嵌入。在構建輔助句的部分，我們提出了兩種方法: AUX-Q和AUX-POSNEG。而堆疊多個嵌入的部分，我們將Transformer-based Embeddings和Static Embeddings堆疊合併成一個新的嵌入。我們在兩個諷刺偵測資料集上實驗我們提出的技術。結果表明，我們提出的方法分別在SemEval 2018 Task 3資料集與News Headlines資料集上，將目前最先進方法的錯誤率降低了7.23%和33.68%。;With the development of social media, the Internet is full of messages. In or der to quickly analyze the positive and negative emotions in a large number of messages, sentiment analysis has become an important issue in natural language processing. Among different types of sentiment analyses, sarcasm detection plays an important role. When a sentence is sarcastic, its meaning will be opposite to that of the core expression. In order to avoid wrong judgments, this research proposes two techniques to improve the accuracy of sarcasm detection: Auxiliary Sentence Construction and Stacking Multiple Embeddings. For the construction of auxiliary sentences, two methods are proposed: AUX-Q and AUX-POSNEG. For stacking multiple embeddings, Transformer-based Embeddings and Static Embeddings are combined into a new embedding. We experimented our proposed techniques on two sarcasm detection datasets. The results showed that our proposed methods reduce the error rate of the state-of-the-art sarcasm detection by 7.23% on SemEval 2018 Task 3 dataset and 33.68% on News Headlines dataset.
