Sentiment Analysis Using Light Weight - Gradient Boosting Machine based Feature Selection

Abstract

Sentiment analysis is a significant task in Natural Language Processing (NLP) that differentiates the emotions and opinions expressed in text or reviews. The sentiment analysis is challenging due to the complex language patterns and inappropriate or redundant features used for classification. In this research, the Light Weight - Gradient Boosting Machine (LWGBM) based feature selection is proposed to choose relevant features for classification to eliminate inappropriate or redundant features and learn the complex language patterns. Then, the classification is performed by using H2O Automatic Machine Learning (H2O Auto ML) algorithm which classifies the sentiments as positive, neutral and negative with high accuracy. The performance of the proposed method is analyzed with different metrics: accuracy, precision, recall and f1-score. The proposed LWGBM and H2O ML method attains an accuracy of 95.39% on the Internet Movie Data Base (IMDB) dataset, and 92.41% accuracy on SemEval - 2016 dataset, which is more effective than the conventional methods namely, Extra-Long Neural Network (XLNet) and Arabic Bidirectional Encoder Representation Transformer (AraBERT).