Sentiment Classification of Arabic Documents: Experiments with multi-type features and ensemble algorithms

Sentiment Classification of Arabic Documents: Experiments with multi-type features and ensemble algorithms Amine Bayoudhi Lamia Hadrich Belguith Hatem Ghorbel ANLP Group, MIRACL ANLP Group, MIRACL ISIC Lab, HE-Arc FSEGS, Sfax University FSEGS, Sfax University Applied Science University 3018, Sfax, TUNISIA 3018, Sfax, TUNISIA CH-2610, Switzerland bayoudhi.amine@gmail.com l.belguith@rnu.fsegs.tn hatem.ghorbel@he-arc.ch Abstract Document sentiment classification is of- ten processed by applying machine learn- ing techniques, in particular supervised learning which consists basically of two major steps: feature extraction and train- ing the learning model. In the literature, most existing researches rely on n-grams as selected features, and on a simple basic classifier as learning model. In the context of our work, we try to improve document classification findings in Ara- bic sentiment analysis by combining dif- ferent types of features such as opinion and discourse features; and by proposing an ensemble-based classifier to investi- gate its contribution in Arabic sentiment classification. Obtained results attained 85.06% in terms of macro-averaged F- measure, and showed that discourse fea- tures have moderately improved F- measure by approximately 3% or 4%. 1 Introduction With the expanding growth of social networks services, user generated content web has emerged from being a simple web space for peo- ple to express their opinions and to share their knowledge, to a high value information source for business companies to discover consumer feedbacks about their products or even to decide future marketing actions. Therefore, opinion mining is becoming a potential research domain interesting more and more researchers who at- tempt to improve current results and to solve more advanced and complex issues in the do- main. Typically, mining opinions is viewed as a classification problem called sentiment classifi- cation. Sentiment classification aims to determine whether the semantic orientation of a text is posi- tive, negative or neutral. It can be tackled at many levels of granularity: expression or phrase level, sentence level, and document level. Expression sentiment classification aims to determine the pri- or sentiment class or valence of an expression. As for sentence level, the objective is to calculate the contextual polarity of a sentence. Concerning document level, which is our focus in this re- search, the main goal is to mine the overall polari- ty of a document with the hypothesis that is ex- pressed by a single author towards a single target. Document sentiment classification is often processed by applying machine learning tech- niques, in particular supervised learning which consists basically of two major steps: feature ex- traction and training the learning model. In the literature, most existing researches rely on n- grams as selected features, and on a simple basic classifier as learning model. The limit of these two choices is revealed when shifting from one domain to another. As a matter of fact, in one hand, each domain has generally his specific vo- cabulary. So, n-grams features produced from one domain fail to be discriminative in another. In the other hand, numerous studies showed that the performance of classification algorithms is domain dependent (Xia et al., 2011). …

Content truncated to 500 words