A Design of a Model for Analyzing Political Sudanese Arabic Tweets Using Deep Learning

Abstract

Sentiment analysis of the Arabic language has gained the attention of many researchers because of the increasing number of Arabic internet users, and the exponential growth of Arabic content online. Sentiment detection of Arabic tweets is at interesting research topic and it enables scholars to analyze huge resources of shared opinions in social media websites such as Facebook and 1 weeter. Sentiment analysis has many applications ranging from financial analysis to political decision-making domains. It is one of the more complex natural language processing tasks due to the informal noisy contents and the rich morphology of Arabic language. Many studies have been investigated for Arabic sentiment analysis. However, most of these works ignore analyzing political Arabic tweets, specifically the Sudanese Arabic dialect. Ehis thesis, focus t: analyz _ Arabic tweets during Sudanese revolution 2018 and developed a deep leaming model for classifying the tweets. The ;w:=;~ "~»cvf:~:»-.?=.~’»~' is a deep learning-based approach. This approach employed word embedding with convolutional neural network and long short-tenn memory network techniques to represent the tweets and extract the feature vectors. Afier that, the feature vectors \\\-to fed to the classifier to detect the sentiments. This approach exploited Word2Vec, convolutional neural networks and recurrent neural networks algorithms. The proposed model has three steps namely data preprocessing, feature extraction, and classification phase. The model has a network architecture to represent the tweets and improve the classification performance. \\. < -_‘.\._1'Li . -»><m~<. » Term Frequency-Inverse Document Frequency and word embedding methods with machine learning classifiers to detect user’s sentiments of colloquial Arabic political tweets instead of using modern standard Arabic. lire»; lllL‘ill\v\lt~ utilized common machine learning algorithms such as \upport \ ector \lachine, Decision trees, ‘~a'i've '--lays, and 'nsemble 1 earning with TF-IDF and word embeddings techniques as feature extraction. Word2vec was used to extract the feature because it is very useful in capturing the syntactic and semantic information of the sentences in the text data. A number of experiments were conducted using a set of performance evaluation metrics on a political lwitter dataset to test the proposed models. Experimental results showed that the proposed deep learning approach outperformed the baseline machine learning methods in tenns of precision, recall and F-score metrics. The proposed _il‘j‘i'\‘.1\']l llll.\ shown .1. ‘~l':‘IllllL ;m1 Hlic in analyzing political sentiment in Arabic tweets. It can be concluded that other Arabic sentiment domains/ applications can benefit from the proposed methods in this thesis.