A Comparative Sentiment Analysis Framework Using Ensemble Learning and Bidirectional GRU Networks on IMDB Reviews

Abstract

Sentiment analysis has proven to be an important Natural Language Processing (NLP) task to understand the user opinions and emotional polarities from long-form texts generated by online platforms and review systems. While traditional machine learning techniques have achieved promising performance in sentiment classification, they suffer from the need for manually designed features and fail to fully exploit the contextual understanding.Traditional machine learning approaches have shown encouraging results for sentiment classification, but they require human-designed features and do not have adequate understanding about the context. To overcome the limitations, this paper proposes a comparative sentiment analysis framework which includes ensemble machine learning models along with deep learning based Bidirectional Gated Recurrent Unit (Bi-GRU) architecture, using the movie review dataset from IMDB. The recommended method consisting of text preprocessing, TF-IDF feature extraction for ML models and word embedding representations for DL application. The classifiers are Decision Tree, AdaBoost, Extra Tree and XGBoost which are evaluated and compared with GRU and Bi-GRU models based on accuracy, precision, recall and F1 score. The experimental results show that the proposed Bi-GRU model achieved the highest classification accuracy (96.4%) than all the traditional machine learning models and the baseline recurrent models. Results show that the framework is effective for contextual sentiment analysis, outperforming unidirectional RNNs in capturing context-sensitive relationships with bidirectional sequence learning, while being scalable and robust to meet the requirements of real-time sentiment analysis.