Improving sentiment analysis in tourism through LLM-enhanced irony detection

As user-generated content continues to proliferate across digital tourism platforms such as TripAdvisor, it has become a critical resource for evaluating service performance and tourist satisfaction within the tourism sector. Among these reviews, ironic ones often reveal underlying issues or dissatisfaction, reflecting more genuine and reliable human emotions to influence users’ choices. However, conventional sentiment analysis methods often fail to recognize irony, which can express negative sentiments through a superficially positive tone. To address the challenge of accurately classifying sentiments in tourism reviews, particularly ironic ones that can mislead traditional sentiment analysis models, this study introduces an innovative sentiment analysis framework named TSI. Leveraging large language models (LLM) for data augmentation, our approach integrates transfer learning and domain-specific knowledge, enhancing detection accuracy through advanced natural language processing (NLP) techniques and deep learning technologies. Extensive experiments on travel-related datasets from Yelp and Tripadvisor demonstrate that TSI outperforms traditional models in both overall sentiment classification accuracy and the identification of negative sentiments, particularly in ironic reviews. This work advances theoretical understanding and practical implementation of sentiment analysis within the tourism sector, equipping businesses with refined tools for interpreting customer sentiments. A new dataset targeting ironic travel reviews is also released to support ongoing scholarly exploration. For better reproduction, the source code is available. 1 1 https://github.com/252820/Tourism-Model .