Sentiment analysis with ensemble and machine learning methods in multi-domain datasets

Abstract

The first place to get ideas on all the activities considered to occur in everyday life was the comments on the websites. This is an area that deals with these interpretations in the natural language processing, which is a sub-branch of artificial intelligence. Sentiment analysis studies, which is a task of natural language processing are carried out to give people an idea and even guide them with such comments. In this study, sentiment analysis was implemented on public user feedback on websites in two different areas. TripAdvisor dataset includes positive or negative user comments about hotels. And Rotten Tomatoes dataset includes positive (fresh) or negative (rotten) user comments about films. Sentiments analysis on datasets have been carried out by using Word2Vec word embedding model, which learns the vector representations of each word containing the positive or negative meaning of the sentences, and the Term Frequency Inverse Document Frequency text representation model with four machine learning methods (Naïve Bayes-NB, Support Vector Machines-SVM, Logistic Regression-LR, K-Nearest Neighbour-kNN) and two ensemble learning methods (Stacking, Majority Voting-MV). Accuracy and F-measure is used as a performance metric experiments. According to the results, Ensemble learning methods have shown better results than single machine learning algorithms. Among the overall approaches, MV outperformed Stacking.