Sentiment Analysis on Bangla and Phonetic Bangla Reviews: A Product Rating Procedure using NLP and Machine Learning

Abstract

The e-commerce industry has expanded quickly in recent years. Online commerce become more well-liked during the reign of Covid. However, maintaining product quality and customer pleasure are significant obstacles for internet businesses. Even when a product has a five-star rating, customers frequently complain about its poor quality. This is so that the quality of the goods can't be specifically described by the star ranking system. The star rating and the actual product reviews don't always agree. Our research study tries to find a solution to this issue. People are more aware than ever before and they try to express their opinion after buying a product by posting reviews on the e-commerce websites. In this study, we attempted to derive rating from customer feedback. The language that is most commonly used in our nation, Bangla and Phonetic Bangla, was our main focus. In order to do this, we first built our own dataset. we gathered over 4,000 Bangla and Phonetic Bangla product reviews from various e-commerce websites, online store pages, social media platforms, and YouTube videos. The data was then divided into two categories, 1 (Positive), and 0 (Negative), and the dataset was preprocessed, which involved cleaning the data and feature extraction. In order to extract features, we employed TF-IDF. Finally, in order to determine the polarity of the reviews, we trained our model using five different supervised machine learning algorithms namely Logistic Regression, Multinomial Naïve Bayes (MNB), Decision Tree, Random Forest and Support Vector Machine (SVM). SVM had the highest accuracy in the Bangla and Phonetic Bangla datasets when the model was tested using test data, and it attained 82% accuracy in Bangla and 94% in Phonetic Bangla dataset.