Detecting Target of Sarcasm using Ensemble Methods

Detecting Target of Sarcasm using Ensemble Methods Pradeesh Parameswaran, Andrew Trotman, Veronica Liesaputra, David Eyers Department of Computer Science University of Otago New Zealand [pradeesh,andrew,veronica,dme]@cs.otago.ac.nz Abstract We describe our methods in trying to detect the target of sarcasm as part of ALTA 2019 shared task. We use combination of ensemble of clas- sifiers and a rule-based system. Our team ob- tained a Dice-Sorensen Coefficient score of 0.37150, which placed 2nd in the public leader- board. Despite no team beating the baseline score for the private dataset, we present our findings and also some of the challenges and future improvements which can be used in or- der to tackle the problem. 1 Introduction We humans are complex creatures that use lan- guage as a communication tool in order to express our thoughts to one another (Sabbagh, 1999). One of the ways that we communicate with another person is through use of verbal irony. Verbal irony is defined as where the words that are being used to communicate differ from the supposed mean- ing (Sperber, 1984). An example of this would be from Austen (1813) Pride & Prejudice, when Darcy said to his future beloved wife, that she is “tolerable but not handsome enough to tempt me”. Sarcasm is a kind of verbal irony that expresses a cynical attitude towards a person or circumstance (Gibbs, 2000). In our daily lives, sarcasm is often conveyed using the tone of our voice, and/or our facial expression, to give the signal to the other person that the person is being sarcastic (Cheang and Pell, 2008). Recently, with the growth of so- cial media many researchers have embarked on various ways of detecting sarcasm automatically. Most of their work were focused on detecting sar- casm on Twitter and on online reviews (Bamman and Smith, 2015; Rajadesingan et al., 2015; Amir et al., 2016). Prior works treated this problem as a binary text classification problem. To the best of our knowl- edge, there is little work that has been done in the realm of identifying the target of sarcasm in sarcastic text. The earliest work in this domain was (Joshi et al., 2019). Identifying the target would help in certain Natural Language Process- ing (NLP) tasks such as in the realm of improving cyberbully detection by helping to identify the tar- get of ridicule (Raisi and Huang, 2016). It has also sparked the organisers at the Australasian Lan- guage Technology Association (ALTA) to organ- ise a shared challenge task to tackle the problem. We employed a 2-phase approach to attempt to solve this task. In our first phase, we employed an ensemble of classifiers along with a meta-classifier to classify sarcasm targets which are marked as “OUTSIDE”. First, we built a Support Vector Ma- chine (SVM) using word embedding to classify the text, followed by the use of a Logistic Clas- sifier. Finally, we used a Linear Classifier to com- bine the results of the two classifiers. In the sec- ond phase of our system, we used …

Content truncated to 500 words