Which Configuration Works Best? An Experimental Study on Supervised Arabic Twitter Sentiment Analysis

Abstract

Arabic Twitter Sentiment Analysis has been gaining a lot of attention lately with supervised approaches being exploited widely. However, to date, there has not been an experimental study that examines how different configurations of the Bag of Words model, text representation scheme, can affect various supervised machine learning methods. The goal of the presented work is to do exactly that. Specifically, this work examines which configurations work best for each of three machine learning approaches that have shown good results when applied on the task of sentiment analysis, namely: Support Vector Machines, Compliment Naïve Bayes, and Multinomial Naïve Bayes. Experimenting with different datasets has shown that each of these classifiers has a Bag of Words configuration in conjunction with which, it consistently performs best. It also showed that some features are dataset dependent.