Building a Sentiment Analysis System Using Automatically Generated Training Dataset

In this paper, we describe a methodology to develop a large training set for sentiment analysis automatically. We extract Arabic tweets and then annotates them for negativeness and positiveness sentiment without human intervention. These annotated tweets are used as a training data set to build our experimental sentiment analysis by using Naive Bayes algorithm and TF-IDF enhancement. The large size of training data for a highly inflected language is necessary to compensate for the sparseness nature of such languages. We present our techniques and explain our experimental system. We use 200 thousand annotated tweets to train our system. The evaluation shows that our sentiment analysis system has high precision and accuracy measures compared to existing ones