Deep Convolutional Network for Arabic Sentiment Analysis

Applying deep Convolutional Neural Networks (CNNs) for Sentiment Analysis (SA) has achieved improvements over the state-of-the-art. CNNs are powerful at extracting hierarchical representation of the input by stacking multiple convolutional and pooling layers. Word embedding is the common approach used for text representation in convolutional networks applied for sentiment analysis. Another technique is combining both word level and character level features. Recently, deep architectures based on character level features only showed more enhanced performance. In this paper two deep CNNs are applied for Arabic sentiment analysis using character level features only. A large scale dataset is constructed from available SA datasets in order to train networks. The dataset maintains opinions from different domains expressed in different Arabic forms (Modern Standard, Dialectal). Besides different machine learning algorithms as Logistic Regression, Support Vector Machine and Naïve Bayes have been applied to assess the performance on such a large dataset. Up to the available knowledge this is the first application of character level deep CNNs for Arabic language sentiment analysis. Results show the ability of Deep CNNs models to classify Arabic opinions depending on character representation only and register 7% enhanced accuracy compared to machine learning classifiers.