The diagram integrates three sections: research motivation, the R o B E R T a C S S process, and research implications. The left section explains that sentiment has a continuous nature and that existing methods only classify discrete categories, leading to oversimplified interpretations. The central section outlines the model workflow in five steps: constructing the data loader, preparing tokens with embeddings, obtaining the classification token output, building the regressor layer, and standardising the output value from 0 to 1. The right section summarises methodological and practical implications. Methodologically, the model offers improved accuracy and reduced cost for continuous sentiment scoring. Practically, it supports tourism departments in identifying satisfaction triggers, allocating resources, and guiding marketing strategies.
The diagram integrates three sections: research motivation, the R o B E R T a C S S process, and research implications. The left section explains that sentiment has a continuous nature and that existing methods only classify discrete categories, leading to oversimplified interpretations. The central section outlines the model workflow in five steps: constructing the data loader, preparing tokens with embeddings, obtaining the classification token output, building the regressor layer, and standardising the output value from 0 to 1. The right section summarises methodological and practical implications. Methodologically, the model offers improved accuracy and reduced cost for continuous sentiment scoring. Practically, it supports tourism departments in identifying satisfaction triggers, allocating resources, and guiding marketing strategies.
Tourist sentiment is typically measured as discrete categories (e.g. positive, neutral and negative) through lexicon-based or machine-learning-based approaches in extant studies. However, neuroscience and physiology scholars have argued that sentiments are continuous in nature. Treating sentiment as a categorical state may result in an overly simplified understanding of tourists’ sentiments, ultimately hindering the tourism industry’s ability to derive precise and actionable insights. This study aims to construct an AI-driven framework for continuous tourist sentiment scoring.
This paper proposed a tool named RoBERTa-CSS (RoBERTa-based Continuous Sentiment Scoring) to calculate tourists’ continuous sentiment scores based on the pre-trained language model RoBERTa. The structure of RoBERTa is refined by adding a fully connected neural network layer, enabling the prediction of continuous sentiment scores. Using Chinese online reviews of a hotel group from multiple travel platforms, 3,500 sentences segmented from 1,000 randomly selected reviews were manually annotated to evaluate the proposed approach.
The comparison with the state-of-the-art open-source packages, deep learning models, pre-trained language models and generative artificial intelligence tools on multiple evaluation metrics demonstrated the superiority of the proposed RoBERTa-CSS. The method was also validated on an English dataset, showing good performance. Several empirical analyses, including individual-level sentiment flow analysis, group-level sentiment distribution and longitudinal analysis, were performed using the full dataset. The results further showcased the edge of RoBERTa-CSS, compared to extant polarity categorization-oriented sentiment analysis methods.
This study expanded the analytical ability beyond simple categorization to facilitate understanding of the complexity and diversity of human sentiment based on an improved pre-trained language model. The relevance of this paper for tourism practitioners, destination management organizations and online travel platforms is discussed.
