Comparative Analysis of BERT-based Classifiers using Semantics of Hotel Reviews

This paper analyses the performance of various classifiers for the extracted fuzzy semantics of hotel reviews available on Tripadvisor. The classifiers rely upon the hidden semantics of reviews which is more insightful representation of user’s thoughts in comparison to ambiguous numerical ratings on a fixed scale. The classifiers are made to learn and predict the user ratings as a function of sentiment score for extracted features. The BERT model is optimized to generate encodings for the classification task. The classification problem is defined as either multiclass or binary classification according to the number of reviews for a particular user rating. The performance of different classifiers namely Logistic Regression, Neural Network, Random Forest, Gradient Boosting and AdaBoost is tested for the BERT based embeddings of the fuzzy semantics derived from the hotel reviews. Further, these classifiers are fine-tuned using hyperparameter tuning to learn the optimal parameters for the prediction of user ratings from the semantics of reviews. The accuracy achieved with single model of logistic regression and other ensemble models is demonstrated in this study. The proposed learning of the classifiers can be utilized to generate muti-criteria recommendations both for users and hotel stakeholders.