Tourism Review Sentiment Analysis for Sustainable Destination Management: Evidence from Four Chinese Tourism-Review Datasets

Online reviews can complement surveys and operational records by providing timely evidence of visitors’ experiences of tourism services, facilities, and attractions. Their value for sustainable destination management, however, depends on reliable interpretation and on their use alongside environmental, social, cultural, economic, and governance evidence. This study evaluates tourism-review sentiment classification as a screening component in such an evidence system. The evaluated label-aware contrastive model jointly contrasts review representations and class-specific representations in text-to-classifier and classifier-to-text directions while retaining a classification objective. We use four Chinese tourism-review datasets covering hotels, attractions, and tourism facilities. On the reported test splits, the evaluated configuration improves on the compared conventional, neural, and pretrained-language-model baselines. The BERT + BiCL configuration attains macro-F1 scores of 90.40, 85.91, 96.15, and 98.82 on the Jiudian, Jingqu, Shanghai, and Beijing datasets, respectively. The three-class Jingqu results also show that the neutral category remains difficult, demonstrating why weighted metrics alone are insufficient for monitoring applications. For decision-making, the model is intended to triage review streams and prioritise records for human examination and comparison with independent evidence; a predicted label neither verifies a service condition nor measures destination sustainability. The study contributes comparative evidence on class-sensitive tourism-review analytics and a bounded interpretation for its use in sustainable destination management.