Multilingual Aspect-based Sentiment Analysis Using Pre-trained Language Models: A Focus on Hotel Reviews

Consumers tend to rely on user reviews when making purchasing decisions, and reviews play a crucial role in corporate marketing and service development. The latest natural language processing technique, ABSA (Aspect-Based Sentiment Analysis), was employed to understand consumer sentiments in multilingual reviews in detail. The training data consisted of Korean, English, and Japanese reviews to detect model performance by linguistic typology. A multilingual pre-trained language model, XLM-RoBERTa(Cross-Lingual Language Model of Robustly optimized BERT pretraining approach), was used to analyze sentiments by aspects. The performance evaluation of the model showed that the model using all three languages—Korean, English, and Japanese—performed the best regardless of linguistic differences. Based on the proposed model, we analyzed the aspects of overseas hotel reviews on a domestic hotel service platform. Through the results of this study, we proposed directions for companies aiming to leverage multilingual reviews in constructing AI training data and implementing ABSA effectively.