Tackling an Unbalanced Dataset for Classifying Indonesian E-Commerce Reviews Using Multi Word Embedding Model

Product reviews are an integral part of ecommerce. By Q2 2020 alone, the average of e-commerce visitors in Indonesia almost reached 400 million. This vast amount of traffic resulting a massive amount of product review information. Product reviews in Indonesia are usually paired with a rating system from one to five based on user satisfaction with the product. Mostly, the number of reviews in each rating is unbalanced. Ratings 5, 4, and 1 are usually more frequent than ratings 2 and 3. This imbalanced data makes classification challenging to categorize the review rating, leading to inaccurate predictions for the rating with fewer data than the others. Previous research used various approaches for classifying review ratings, such as machine learning and fine-tuning a state-of-the-art deep learning model. In this research, we propose a new approach that leverages multiple-word embedding combined with CNN to deal with the highly unbalanced dataset problem. By leveraging multiple-word embedding techniques, the model can extract more features from the review text. We also proposed an extended pipeline to enhance the model performance. The proposed model and pipeline perform better than the fine-tuned Indonesian BERT-based and RoBERTa-based pre-trained models baseline with 0.93 accuracy and 0.82 F1-Macro scores.