Combing Semantic and Prior Polarity Features for Boosting Twitter Sentiment Analysis Using Ensemble Learning

Twitter sentiment analysis provides organizations with real-time monitoring of public feelings towards particular products and events related to them. Most existing research is focused on extraction of sentiment features through analysis of lexical and syntactic features that are expressed explicitly through words, emoticons, exclamation marks etc. Single machine learning classifiers are usually employed by these approaches for tweet sentiment classification. In this paper, we introduce a semantic feature model that utilizes co-occurrence statistics and latent contextual semantic relationships between words in tweets. These semantic features are combined with prior polarity score features and n-grams features as a sentiment feature set of tweets. The feature set is incorporated into an ensemble classifier formed by Support Vector Machines, Logistic Regression and Random Forest for training and predicting sentiment classification labels. Five Twitter datasets are used in our evaluation and the performance of the model is compared with a word n-grams model as a baseline. Experimental results show that our method has superior performance in terms of the accuracy of the sentiment classification and that ensemble learning can be used to improve classification accuracy.