A Transformer Based Approach for Analyzing Scrapped E-Commerce Product Reviews from Social Media Platforms with Explainable AI

Online shopping has gained substantial popularity in Bangladesh, where companies need to understand the sentiments of customers in order to enhance their products and services for their business. Furthermore, it assists customers to make better purchasing decisions. With the advance of Bengali Natural Language Processing (BNLP), numerous Machine Learning algorithms have performed significantly in classifying sentiments from Bengali sequential data. Cultural and dialect variations limit the capacity of these architectures, where most Recurrent Neural Network-based architectures cannot perform better in classifying Bengali data. Bangla BERT architecture is a transformer-based architecture that can perform significantly better in Bengali sentiment analysis. In this research, the authors gather a dataset of 5016 comments associated with product ID, rating, and sentiment towards a particular product. Exploratory Data Analysis reveals that products with poor ratings exhibit good comments from the same customer. Hence, the Bangla BERT model is applied to the customer’s comments from the dataset to achieve better performance with necessary preprocessing and hyperparameter tuning. The experimental result shows that the model reflects a 95.02% F1-score that outperforms state-of-the-art architectures. Finally, an Explainable AI model named LIME is applied to understand the feature extraction procedure of the transformer model.