Sentiment Analysis of Tunisian Users on Social Networks: Overcoming the Challenge of Multilingual Comments in the Tunisian Dialect

Abstract

The presence of the dialect in the Arabic texts made Arabic sentiment analysis (ASA) a challenging issue, owing to it usually does not follow specific rules in writing systems, especially Tunisian Dialectical (TD) which presents an undertaking challenge due to its complexity, ambiguity, the morphological richness of the language, the absence of contextual information, the code-switching (CS) and mostly the multilingualism phenomena in textual productions. Recently, deep learning models have clearly demonstrated a great success in the field of sentiment analysis (SA). Although, the state-of-the-art accuracy for dialectical sentiment analysis (DSA) still needs improvements regarding contextual information and implicit sentiment expressed in different real cases. To address this challenge, we propose, an efficient Bidirectional LSTM network preceded by a preprocessing stage in order to enhance Tunisian SA, by applying Forward-Backward encapsulate contextual information from multilingual feature sequences. To evaluate our model, and due to the lack of publicly available multilingual resources associated with the TD, we collect different datasets available with different variants of TD to create our own multilingual corpus for sentiment classification. The experimental results based on the evaluation standards “Accuracy”, “Recall” and “F1-score” demonstrate that our model achieves significant improvements over the state-of-art deep learning models and the baseline traditional machine learning methods.