Sentiment Classification of Sinhala Content in Social Media: An Ensemble Approach

Abstract

We focus on the binary classification of Sinhala social media content in the sports domain using machine learning algorithms. In particular, we improve upon the accuracy achieved in a previous study of ours that utilized word and character N-grams. We use the base learners from that study to implement a probability-based stacking ensemble approach. This is done by creating a base learner library of 1066 base learners, using 13 different algorithms and different N-gram feature extraction methods. Different base learner combinations from the library are then stacked together to find the best stacking ensemble model. The best stacking ensemble model achieves an accuracy of 83.8% which is an improvement of over 1.5% of our previous study.