Abstract
This study is based on a benchmark corpora consisting of 3,015 textual Arabic opinions collected from Facebook. These collected Arabic opinions are distributed equally among three domains (Food, Sport, and Weather), to create a balanced benchmark corpus. To accomplish this study ten Arabic lexicons were constructed manually, and a new tool called Arabic Opinions Polarity Identification (AOPI) is designed and implemented to identify the polarity of the collected Arabic opinions using the constructed lexicons. Furthermore, this study includes a comparison between the constructed tool and two free online sentiment analysis tools (SocialMention and SentiStrength) that support the Arabic language. The effect of stemming on the accuracy of these tools is tested in this study. The evaluation results using machine learning classifiers show that AOPI is more effective than the other two free online sentiment analysis tools using a stemmed dataset.