Multi-Label Classification of Emotions in Arabic Tweets From Different Perspectives

Sentiment analysis has been studied widely in the literature. Despite these many works, there is a lack of works that concentrates on finding the basic emotions behind these sentiments. This problem becomes more challenging in under-resourced languages, such as Arabic. Furthermore, to our knowl-edge, no works have studied this problem from the reader's versus the writer's perspectives. In this work, we study sentiment analysis and basic emotions extraction from the writer's perspec-tive, which is the person who wrote the text, to complement an earlier work of ours focusing on the reader perceptive. Using a dataset of Arabic tweets, we compare the two perspectives. Since each tweet may contain multiple emotions, we use Multi-Label Classification (MLC) techniques. Three classifiers are compared: Decision Trees (DT), Random Forest (RF), and K-Nearest Neighbor (KNN). Prior results showed that the top performing classifier in the case of the reader dataset was RF. For this work focusing on the writer dataset, RF is not a clear winner for all performance metrics under consideration as DT produces competitive results.