Enhanced Sentiment Analysis in Microblogs through the use of XGboost Classifier and Genetic Algorithm

Studies on sentiment analysis and opinion mining initially focused on polarity classification through the use of positive, negative, or neutral categories. Nevertheless, despite their importance in a wide range of applications, the classification of extreme opinions, such as highly negative and very positive ones were not targeted until recently. In this work, we focus on a 5-point scale to include extreme sentiments as well. The majority of studies in this domain have focused on approaches tailored towards special datasets. This doctoral thesis proposes two novel ensemble classifier approaches to improve the performance of the sentiment analysis task. The first proposed ensemble classifier framework called “SentiXGboost” is designed to improve binary sentiment analysis tasks using the XGBoost algorithm as a meta-classifier for stacked ensembling. The second proposed approach provides a framework based on the concept of the Genetic Algorithms for producing an optimized classifier ensemble for binary, ternary, and fine-grained, denoted “SentiGA”, sentiment analysis task. Both of the proposed approaches are evaluated on the major sentiment datasets, including SemEval-2017 (Sentiment Analysis in Twitter) task (4A, 4B, and 4C), Stanford Sentiment Treebank (SST-2 and SST-5), Sentimet140, Sentiment Labelled Sentences (Amazon), Stanford Sentiment Gold Standard, Yelp Challenge Dataset and Movie Review (Sentiment Polarity Dataset V2.0). The performance of both proposed approaches is compared with other existing well-known methods in the field using the same datasets. The results show that our proposed approaches have successfully enhanced the performance of sentiment analysis classification compared to other existing methods.