Multi-class Sentiment Classification for Customers’ Reviews

The rise of e-commerce due to the Covid-19 situation is becoming more significant in 2021. It could lead to great demands to understand customers' opinions usually shown in their reviews. An e-commerce platform with the ability to be aware of its users' viewpoint can have a higher possibility of meeting customer expectations, attracting new users, and increasing sales. With the tremendous data in e-commerce platforms presently, sentiment analysis is a powerful tool to understand users. However, the sentiment in reviews data may contain more than two states, positive and negative, and then a binary sentiment classifier may not be helpful in practice. According to our knowledge, research on this subject is often restricted access. Therefore, this paper presents a multi-class sentiment analysis for Vietnamese reviews on a large-scale dataset, including 480,702 reviews. We collected these reviews from popular Vietnamese e-commerce websites and manually did the labeling process with three classes of sentiments (positive, negative, and neutral). To build a suitable classification model for the main problem, we propose a deep learning approach using different architectures (LSMT, GRU, TextCNN, LSTM + CNN, and GRU+CNN) and compare the performance among other ensemble techniques. The experimental results show the outperformance of the ensemble techniques on the multi-class sentiment classification problem, and the combination of chosen architectures using the attention mechanism could obtain the best F-1 score of 73.64%.