A Cantonese Restaurant Review Dataset for Aspect Category Sentiment Analysis with XLM-RoBERTa

Abstract

The volume of online consumer reviews has surged as e-commerce continues to grow, providing valuable insights for both consumers and business owners. Aspect Category Based Sentiment Analysis (ACSA) identifies sentiment polarity based on specific aspect categories in reviews. Despite extensive research in other languages, Cantonese remains underexplored due to its unique linguistic features and limited datasets. This study seeks to bridge this gap by fine-tuning a RoBERTa-based model for Aspect Category Sentiment Analysis (ACSA) on Cantonese restaurant reviews from OpenRice. A dataset is constructed by collecting and annotating 7,473 reviews based on five aspect categories: food, service, ambience, price, and timeliness. The fine-tuned XLM-RoBERTa model achieved an accuracy of 75.17%, outperforming five baseline models and demonstrating the efficacy of transformer-based models in low-resource languages. The study shows that the fine-tuned RoBERTa-based model has significant advantages in processing low-resource languages such as Cantonese, not only surpassing the baseline model in accuracy, but also providing a solid foundation for future research on Cantonese sentiment analysis. This work contributes a significant dataset and highlights potential future research directions.