ArSEL: A Large Scale Arabic Sentiment and Emotion Lexicon

With the advancement of Web 2.0, social networks experienced a great increase in the number of active users reaching 2 billion active users on Facebook at the end of 2017. Consequently, the size of text data on the Internet increased tremendously. This textual data is rich in knowledge, which attracted many data scientists as well as computational linguists to develop resources and models to automatically process the data and extract useful information. One major interest is sentiment and emotion classification from text. In fact, learning the opinion and emotions of people is important for businesses, marketers, government, politicians, etc. While focus had been given to sentiment analysis, recently emotion analysis has captured great interest as well. Several resources were developed for emotion analysis from text for English, however, very few targeted Arabic text. We present in this paper, ArSEL, the first large scale Arabic Sentiment and Emotion Lexicon. ArSEL is built in a way to augment the publicly available Arabic Sentiment Lexicon, ArSenL, and to generate a large scale lexicon that includes emotion and sentiment labels for almost every lemma in ArSenL. We also show the efficiency of using ArSEL in emotion regression and classification tasks using an Arabic translated version of annotated data from SemEval 2007 “Affective Task” as well as SemEval 2018 Task1 “Affect in Tweets” Arabic dataset. Coverages of 91% and 84% are achieved on the two datasets respectively. An improvement of 30% compared to majority baseline is achieved in terms of average F1 measure for emotion classification on SemEval 2018 Arabic dataset. ArSEL is publicly available on http://oma-project.com.