Sarcasm Detection Using Feature-Variant Learning Models

Abstract

Sentiment Analysis is the text classification tool that analyses a sentiment, message, emotion, attitude and tells whether the sentiment is positive, negative or neutral. The prime challenging aspect of sentiment analysis is the presence of sarcasm in message. Sarcasm is one kind of sentiment that is expressed verbally through the use of rolling of eyes and tonal stress. It consist of words mean the opposite of what user want to convey in order to be funny, or to show some irritation. The active online users and their reviews on websites are large in number so it is hard to detect even for humans, so in order to achieve error-free sentiment analysis it is imperative for machines to detect it accurately. The paper proposes the use of three different classes of features to help computers identify sarcasm reasonably well. In this paper, we intend to implement and empirically analyze number of computing techniques like Support Vector Machine, Decision Trees, Logistic Regression, Random Forest, K-Nearest Neighbors and Neural Networks for sarcasm detection on social media. The experimentation was done using three datasets i.e. SemEval 2015 Twitter benchmark dataset; random tweets collected using the Streaming API and a publicly available dataset of Reddit posts. The datasets provide interesting insights into how different forms of social media use the tool of sarcasm differently. The evaluated results were based on the performance measures like precision, recall, accuracy and F score. Amongst all, Twitter datasets had achieved the highest accuracy of around 91% to 92%, while the Reddit dataset had obtained peak accuracy of 80%.