Opinion Mining Using LSTM Networks Ensemble for Multi-class Sentiment Analysis in E-commerce
Dima Alnahas, Furkan Asik, Alper Kanturvardar, Abdullah Mert Ülkgün · 2022
As technology conquers more sectors and industries, the burden of protruding falls heavy on stakeholders, especially with the rising popularity of blogging and social media. The outpour of views and opinions, accompanied by the recent trend of influencers and content creators, inspired various advances in natural language processing. These advances aim to harvest reviews and experiences of consumers to start, improve, and expand businesses. The recent surge in textual data has also inspired many large-scale data analyses powered by artificial intelligence. In this study, we investigate some of these analyses and their applications. We also implement state-of-the-art LSTM models to explore different attributes of text. Based on the relevancy of these attributes, we suggest an ensemble model of LSTM neural networks that utilize word-level and character-level tokenizers to convey helpful knowledge in sentiment or opinion classification. For training, validating, and testing the presented model, we accumulated a large dataset of 602,202 product reviews from Turkish e-commerce websites. The study establishes an 84.7% accuracy of the proposed ensemble model when applied to raw data.