Ensembles for Text-Based Sarcasm Detection

Abstract

Sarcasm is a big challenge for text related classification such as sentiment analysis and opinion summarization. The nature of sarcasm is to express opinions in a way that carries the opposite sentiment as a way to insult or make fun of the situation. Because of its nature, it poses a very difficult challenge in text classification task as they will be classified according to words used and not the meaning implied. Therefore, the accuracy of classification will be affected significantly. Sarcasm is also a problem for the task of sentiment analysis and emotion detection, as they reflect opposite sentiments of the author. So, sarcasm detection is needed to find the sarcasm text and revert the sentiment of the text. In the recent works seen in sarcasm detection, machine learning methods and deep learning methods are more commonly used to perform the task. Although deep learners are efficient learners, machine learner are still widely used and can perform as well as deep learners with proper training. This work seek to compare the different ensemble settings to evaluate the performance of ensembles against simple learners. The results show that ensembles can improve the performance of simple learners and even deep learners.