Classifying sentiment in arabic social networks: Naïve search versus Naïve bayes

Abstract

Social networks contain large amounts of posts of different data types (text, images, sounds and videos). Textual posts express authors' opinions (with or against) or feeling (love, hate, optimism, pessimism, or anger). Such opinions are important for commercial and governmental organization since they help checking public opinion about a product, policy or an object in general. In this paper we present the application of two different approaches to classify Arabic Facebook posts. The first one depends on syntactic features, using common patterns used in different Arabic dialects to express opinions. These patterns achieved high accuracy in determining the polarity of a sentiment even when tested against new corpus. This approach acts on informal Arabic text, which has not been addressed before. Different setups were tried and the highest coverage and accuracy achieved were 49.5% and 83.4 % respectively. The second approach is an ordinary probabilistic model, Naïve-Bayes classifier, that assumes the independence of features in determining the class the highest coverage achieved in this approach was 60.5% in the first setup and 91.2% when Naïve search was used as a binary classifier to classify the posts as objective or subjective.