Sentiment Analysis Of Arabic Tweets Based On Ensemble Machine Learning Approach

Abstract

Sentiment analysis is a powerful technique used to analyze and classify opinions and emotions expressed in textual writings. The increased use of social networks in the Arab world has provided a rich research ground for sentiment analysis. In this thesis, the problem of sentiment classification of Arabic text was addressed. A model based on machine learning was proposed to classify the underlying sentiments as being positive or negative. The proposed model is a voting ensemble of three classifiers: Support Vector Machine, AdaBoost and Naive Bayes Multinomial. To evaluate the effectiveness of the proposed model, a dataset was created from tweets discussing several polarizing social events associated with Saudi Arabia’s vision 2030. The dataset was manually annotated according to sentiment. In addition to the created dataset, an earlier dataset was used in this thesis as a baseline dataset. The proposed sentiment analysis framework involved several linguistic techniques, including light stemming and word normalization. Also, multiple feature sets were extracted and explored such as N-grams, tweet topic and emoticons-based features. Furthermore, the obtained results were optimized by finding the optimal classifier parameters, investigating several ensemble combination rules and applying correlation-based recursive feature elimination. The experimental results revealed excellent classification accuracy and portrayed the ensemble’s ability to perform better than the performance of its base classifiers. Moreover, the results indicated that the proposed ensemble technique outperformed the earlier used model on the same baseline dataset by an 8.5% accuracy increase, which is an indication of the superiority of the proposed model.