Thesis (Ph.D.) - Cairo University - Institute of Statistical Studies and Research - Department of Computer and Information Sciences

Recently, there is a large amount of text provided in different forms such as, forums, twitter messages, and online news websites. Unlike studies that aim to classify text content into one of the binary classes {positive , negative}, there is a great interest recently in one of opinion mining tasks that aims to extract emotions from the text content, this task studied widely in different languages such as Chinese, English or Spanish language, and few attempts are concerned about Arabic text content. It is now widely acknowledged that social media applications did play a role in the Arab spring uprising especially, Egyptian's uprising in 2011 and people use them to spread out their opinions and their thoughts. In this thesis, we are contributing in the research area of mining these political opinions by detecting five fine-grained emotion categories {happiness, sadness, fear, anger, disgust} and mixed emotions {happiness/sadness, sadness/fear, anger/disgust} from Arabic/Egyptian colloquial microblog text in twitter. To achieve that, we applied three different approaches; lexicon based approach, a combined lexicon based approach and Multi-Criteria Decision Making approach and finally, machine learning approach. The dataset consists of 1477 tweets in the political domain, as a reflection to the most popular political issues raised after 25th of January 2011: #أزمة_السولار، #25jan, #اداء_الحكومة، #30June, #anti_coup, #أزمة_الكهربا. Results reveal that lexicon based approach achieved 78.81%. The combined approach achieved promising results as the successful classified tweets are analytical represented by Co-Plot in order to add a visual enhancement representation for emotion intensities calculated by the proposed emotion scoring algorithm. Finally, in machine learning based approach, the ten-fold cross validation for Naïve Bayes classifier and Support Vector Machine classifier achieved 87.81% and 88.96% respectively. While the results are promising, there are still challenges to overcome such as the limitation of emotion words and ambiguity

Issued also as CD