Comparative Evaluation of Unigram and Bigram Models for Sentiment Classification of Hotel Reviews

This paper provides a comparative analysis of as much as possible piece of N-gram-based sentiment classification models to huge-scale hotel review data. Three feature extraction methods such as unigram, a bigram, and a combination of both unigram and bigram representations were used to examine sentiment polarity with respect to negative, neutral and positive classes. Machine learning classifiers (Logistic Regression, Support Vector Machine (SVM), Naive Bayes, Ridge, Stochastic Gradient Descent (SGD) and Passive Aggressive) were tested on three baseline models (M1-M3). An experiment shows that the highest performance is achieved by the unigram-bigram model (M3), and the accuracy, precision, and recall indicate that M3-UniBigram-SVM had an evaluation of 0.8538, 0.8298 and 0.8538 respectively. The analysis of confusion matrices revealed that the highest accuracy was obtained when positive sentiment was considered and the most difficult category according to semantic ambiguity was that of neutral sentiment. The findings prove the hypothesis that hybrid N-gram representations can greatly improve the sentiment classification performances when compared to either single unigram or bigram models. The research generates potent baseline models that are used to support highly developed hybrid sentiment classification models.