Arabic opinion target extraction from tweets

Abstract

Twitter is an ocean of sentiments; users can express their opinion freely on a wide variety of topics. The unique characteristics that twitter holds introduce a different level of challenge in the field of sentiment analysis. Identifying the topic or the target of the expressed opinion is the aim of this study; Opinion target recognition is a task that has not been considered yet in Arabic Language. In this paper we propose a method to extract the opinion target from tweets written in Arabic language. The task is carried out in three phases. Phase 1: preprocess the tweet to delete unnecessary entities like mentions and URLs. Phase 2: construct a feature set from tweet words to be used in the classifying phase; these features are part-of-speech, Named entities, English words, tweet hash tags and part-of-speech pattern. Phase 3: Three classifiers are trained using the extracted features, to assign each word in the tweet to be either an opinion target or not, these classifiers are: Naïve Bayes, Support vector machine and k-nearest neighbor, with an F-Measure result reaching 91%. 500 tweets are used for the experiment, where the opinion target was manually tagged. Finally, a comparison between the results of each model is conducted.