Emoji-Based Sentiment Analysis of Arabic Microblogs Using Machine Learning

Nowadays, there is an explosive growth of social data analytics to measure human reactions to various events. However, the application of natural language processing for text analysis and mining of microblogs is a daunting task with an excessive complexity. In this paper, we explore the idea of adopting new non-verbal features for sentiment analysis of microblogs. We considered 969 emojis and prepared a dataset of 2091 instances written in multidialectal Arabic and each contains at least one emoji. Several machine learning algorithms are evaluated on the suggested features. The experimental results demonstrated that emoji-based features alone can be a very effective means for detecting sentiment polarity with high performance. For instance, when using a multinomial naive Bayes classifier, an F ₁ score of 80.30% and AUC of 87.30% were achieved using the 250 most relevant emojis.