Bangla E-Commerce Sentiment Analysis: Optimization using Tokenization and TF-IDF

Abstract: This paper aims to explore the sentiments of Bangla customer reviews and provide valuable insights for potential buyers. The proposed methodology involves a comprehensive approach to data collection from online shopping platforms, carefully removing duplicates and irrelevant information and labeling sentiments. Trigram features were used in combination with TF-IDF for effective feature selection. Data augmentation was also utilized to address the class imbalance. Various classifiers were deployed to categorize sentiments, including Naive Bayes, Random Forest, Extra Tree, KNN, Logistic Regression, Support Vector Machine, Artificial Neural Network, and Multilayer Perceptron. In analyzing BangIa product sentiments, the Multilayer Perceptron model performed exceptionally well and achieved an accuracy rate of 96.42% using the proposed methodology. The Artificial Neural Network closely followed it with 96.23% and Naive Bayes with 95.38%, providing valuable insights into the BangIa customer sentiment analysis landscape.

• Comprehensive Data Collection: Data gathered from online shopping platforms, removing duplicates and irrelevant info.
• Sentiment Analysis Focus: Explores Bangla customer review sentiments for potential buyers.
• Sentiment Labeling: Reviews labeled accurately to identify sentiments.
• Feature Selection: Utilized trigram features with TF-IDF for effective feature selection.
• Data Augmentation: Addressed class imbalance through data augmentation.
• Classifier Deployment: Used Naive Bayes, Random Forest, Extra Tree, KNN, Logistic Regression, SVM, ANN, and MLP classifiers.
• High Accuracy Rates: MLP achieved 96.42%, ANN 96.23%, and Naive Bayes 95.38%.