Aspect-based Sentiment Analysis for Vietnamese Reviews about Beauty Product on E-commerce Websites

Aspect-based Sentiment Analysis for Vietnamese Reviews about Beauty Product on E-commerce Websites Quang-Linh Tran, Phan Thanh Dat Le, Trong-Hop Do University of Information Technology, Ho Chi Minh City, Vietnam Vietnam National University, Ho Chi Minh City, Vietnam {18520997, 18520570}@gm.uit.edu.vn, hopdt@uit.edu.vn Abstract Millions of reviews are generated on e- commerce platforms and analyzing them deeply brings a lot of useful information for sellers and buyers. This paper deals with the aspect-based sentiment analysis problem to an- alyze the aspect and polarity of Vietnamese reviews about beauty products on e-commerce websites. The contribution of this paper is three-fold. Firstly, we introduce a dataset con- taining 16,227 reviews about lipsticks. There are 32,775 pairs of aspects and sentiments in this dataset. Besides, we conduct some baseline experiments using Deep Learning-based mod- els to build detection and classification models for extracting the aspects of reviews and classi- fying the sentiment of each aspect. In addition, a comprehensive comparison is also performed to see whether single-task learning or multi- task learning is the better approach to predict the aspect and sentiment of reviews. The exper- imental results show that the BiGRU+Conv1D model in the single-task learning approach out- performs others with the F1-score in the aspect detection task at 98.09% and 91.01% for the sentiment classification task. 1 Introduction Aspect-based sentiment analysis (ABSA) is a task in Natural Language Processing (NLP). Instead of only extracting the polarity of the text in sentiment analysis, ABSA analyzes the sentiment more deeply by showing the polarity of each aspect of the text. In reviews on e-commerce platforms, a review con- tains more than one aspect so ABSA is more robust to analyze the polarity than sentiment analysis. For example, in lipstick reviews, the customers do not always give general reviews about the products, they tend to review based on the aspect of the lipstick such as price, colour, et cetera. The customers may like the colour, but the texture of the lipstick is so bad that they are disappointed with this. This is a reason why ABSA is really necessary to understand more deeply what customers think about the products. With the explosion of e-commerce, millions of re- views are generated every day from customers. They leave a lot of useful information for not only the sell- ers to understand what their customers like and dis- like about their products but also for other customers to read and consider previous buyers’ experiences before making a purchase. There are many kinds of products in e-commerce, but beauty products are one of the most favorable products to buy online. In addi- tion, the reviews of beauty products, especially about the lipsticks contain a lot of aspects, from price, stay- ing power to colour, thanks to which, several analyses can be conducted to deeply understand the attitude of customers toward the lipstick. This is the reason why we choose reviews about lipsticks to build a dataset to deal with ABSA problem in e-commerce reviews. There are three main contributions of this …

Content truncated to 500 words