Improving Sentiment Classification for Large-Scale Social Reviews Using Stack Generalization

The problem of identifying sentiment from customers’ reviews has been an important issue for many years. Previously, different machine learning methods have been utilized to automatically categorize users’ reviews into polarity levels such as positive, negative, or neutral. However, these methods suffer from low accuracy and recall. This paper presents an ensemble learning method using stacking generalization to build an accurate model for predicting sentiment polarity from social reviews. The basic concept of stacked generalization is fusing the output of a first-level classifier with a second-level classifier in a stacking manner. The diversity among the base classifiers with different features and weight measures is investigated in two domains (Twitter and Amazon product reviews), which provides a space for improving sentiment classification performance. Four types of singular classifiers: namely, support vector machine, boosted decision tree, Bayes point machine, and averaged perceptron, are used to build a two-staged and stacking model. The performance of singular and two-staged classifiers is compared with the proposed stacking model. The experiment results demonstrate that the stacking model outperforms the singular and two-staged classifiers on both datasets in terms of accuracy, precision, recall, and F1-score.