Multilevel Sentiment Analysis in Arabic

Abstract

Sentiment analysis has a great necessity to classify sentences like review, news, blog, etc. in order to hold the overall sentiment (i.e. negative, positive or neutral) embedded in them. The vast majority of studies focused on sentiment analysis for English texts, while there is small number of researches has focused on other texts such as Arabic, Turkish, Spanish and Dutch. In this study, we aimed at improving the performance results of Arabic sentiment analysis in the level of document by: firstly, investigating the most successfully Machine Learning (ML) methods to classify sentiments, at the same time rules have been implemented to create new vector formats for representation of inputs with ML based modeling process. Secondly, applying Lexicon Based (LB) approach in both term and document levels by using different formulae based on aggregating functions like maximum, average and subtraction. However, the rules have been applied in the experiments. Performance results of LB approach have been used to identify the best formulae can be used with term level and document level of lexicon based SA at Arabic Language, also the effectiveness of using rules in both levels has been illustrated. As a final point, employed methods of the two different approaches (i.e. ML and LB) have been tried to create a combined method with considering rules. The OCA corpus was used in the experiments and a sentiment lexicon for Arabic sentiments (ArSenL) was used to resolve the challenges of Arabic Language. Several experiments have been performed as followed: Firstly, features have been selected for both term and document levels of the OCA corpus independently. Secondly, different linear ML methods such as Decision Tree (D-Tree), Support Vector Machine (SVM), and Artificial Neural Network (ANN) have been applied on both of OCA corpus levels with considering applying and not applying rules on both levels of the corpus. Thirdly, LB approach have been applied on the document level with considering applying rules to each term in a document. And finally comparisons between the results have been done to identify the best way to classify sentiment Arabic documents. The most successful results in the study are as follows: (i) In ML approach, ANN classifier has been nominated as best classifier in the term level and in the document level of Arabic SA. Furthermore, the average of F-score achieved in the term level for positive testing classes is 0.92, and also in negative classes is 0.92, however, in the document level, the average of F-score for positive testing classes is 0.94, while in negative classes is 0.93. (ii) In the LB approach, it is concluded that the best results have been achieved by applying rules for each term, then computing each sentence score by DMax_Sub formula, and finally, using first sentence score formulae for document score computing. In general, the results of the ML approach are better than the results of the LB approach.