Sentiment analysis of online reviews on hotel booking websites using transformer models

Abstract

Online reviews or comments about products or services in online media formats, such as reviews and comments from individuals who have purchased the product, are a type of Electronic Word of Mouth (eWOM) that allows consumers to search for relevant information before making purchasing or using decisions. The volume of review and commentary information today is enormous and constantly emerging, requiring consumers to analyze the pros and cons of these products before making decisions. This study collected hotel review data from Agoda.com, which had over 100,000 reviews, to analyze the feelings toward the hotel business. The data was processed using Transformer-based models, including Multilingual BERT, XLM-RoBERTa, and WangchangBERTa, as well as original feature extraction techniques such as Bag of Words and TF-IDF. Then, the researcher used Logistic Regression to analyze the reviews. In the experimental result, each model predicted the correct outcome, with XLM-RoBERTa having the highest accuracy rate of 94.55%, followed by WangchangBERTa with 91.45%, Multilingual BERT with 90.40%, and TF-IDF and Bag of Words with 91.45% and 91.15%, respectively. The XLM-RoBERTa resulted in the highest accuracy rate for predicting sentiment analysis results through deep learning, making it suitable for building models for analyzing Thai text. Implication of thematic paper: The findings of the study indicate that transformer-based models offer significant advantages and efficiencies for sentiment analysis by accurately analyzing customer sentiment data. The insightful data could support businesses in terms of decision-making, improving customer experiences, and optimizing marketing strategies.