Enhancing the Performance of Sentiment Analysis by Using Different Feature Combinations

With the growing accessibility and acceptance of social networking epoch, Sentiment Analysis has become one of the most prominent research domain in natural language processing. Each day, millions of people share their thoughts and ideas by posting in social media or writing online reviews. This massive participation, on one hand, makes these media opinion-rich; however, on the other hand, it poses some challenges in identifying the dominant opinion. In this work, tweets and movie reviews are classified according to the polarity of the opinions by using several features in combination. Performance of several feature combinations was evaluated by feeding those in different Machine Learning algorithms (NB, SVM, MaxEnt). Hence, the goal of the work was to evaluate how the performance of a classifier is affected when different feature combinations are used in Sentiment Analysis. Experiments were done on data from two different domain namely Stanford Twitter Sentiment 140 dataset and IMDb Movie Reviews dataset. Four different evaluation metrics: recall, precision, accuracy and F1 score are used for evaluating the investigational results of our system. This research demonstrates that by carefully choosing correct feature combination the classification accuracy can be increased while a random feature combination will provide little benefit.