Semi-Automatic Category Estimation and Data Augmentation for Opinion Extraction of Product Components

Abstract

When customers purchase a product online, they use reviews to gather information about that product to help them make a purchase decision. Aspect-based Sentiment Analysis is a task that analyzes the review content from various perspectives, including the product itself, its components, and its retail outlets. We focus on comparing the characteristics of each component in a product with those of other products at the time of purchase. We define a task called component-based sentiment analysis (CBSA), which analyzes the review content from the perspective of only each component in the product. The CBSA task consists of opinion target extraction and polarity analysis. We approach that task with a classifier. We describe a semi-automatic category determination method for creating classification labels for CBSA and a data augmentation method to improve its classification performance. In experiments, we show that our category determination method can generate categories that cover 95% of the existing categories on e-commerce sites and that our data augmentation method improves the macro-F1-measure for uncommon opinions by 10%.