Sentiment Analysis of Review Data: A Deep Learning Approach Using User-Generated Content

Abstract

As information technology progresses rapidly, social media platforms are experiencing exponential growth, accompanied by a surge in online content. Sentiment analysis (SA) of online evaluations has piqued the interest of researchers from various organizations, including academia, government, and private industry. It has become an increasingly hot area of study in the fields of Machine Learning (ML) and natural language processing (NLP). Deep Learning (DL) algorithms are currently being utilized in the same field to achieve remarkable results. Much SA research has been conducted in different languages such as English, Chinese, and Spanish, as well as various Indian languages like Hindi, Malayalam, and Bengali. However, languages like Assamese have received very little attention in this field of research. Hence, this research work provides a novel approach to sentiment analysis by demonstrating the effectiveness of deep neural network models for the less explored and scarce resource language, Assamese. This paper introduces a hybrid model, combining Convolutional Neural Network (CNN) and Long Short-Term Memory (LSTM), termed LSTM-CNN, for Sentiment Analysis (SA). Keras word embeddings are employed for vectorization of the data. To achieve outcomes, our proposed models have employed dropout, max pooling, and batch normalization techniques. Experimental analysis is carried out on user-generated review content by translating the Bengali dataset into Assamese. Comparison and evaluation of the built models have been done with traditional machine learning models in predicting sentiment polarity. Comparative analysis shows that all the proposed models outperform with an accuracy of more than 98%.