This study aims to develop a robust analytical framework for identifying and interpreting negative user-generated reviews of five-star hotels. It addresses the limitations of traditional review filtering based solely on star ratings and the manual subjectivity of topic modeling, offering a deep learning–based solution aligned with national hospitality standards.
Using a data set of 124,381 user reviews from 70 five-star hotels in Jiangsu Province, collected via the Ctrip platform, this study applies a fine-tuned Chinese bidirectional encoder representations from transformers (BERT) model to detect negative reviews with high semantic accuracy. BERTopic is then used for topic modeling. To enhance domain relevance and interpretability, a high-quality semantic vocabulary is constructed based on the national standard Classification and Accreditation for Star-Rated Tourist Hotels (GB/T 14308–2023). The extracted topics are mapped to this vocabulary to establish structured semantic alignment between customer feedback and industry evaluation dimensions.
The BERT model identified 18,578 negative reviews, a figure significantly exceeding the number captured by rating-based filters alone. Among these, 437 topic clusters were extracted via BERTopic, with 388 successfully mapped to a standardized topic vocabulary. Results highlight that negative feedback is concentrated in key service areas such as room facilities, cleanliness, staff responsiveness and safety assurance. Notably, approximately 13% of high-rated (4–5 stars) reviews also contained negative sentiment, exposing service blind spots hidden beneath favorable scores.
This study focuses on Chinese-language five-star hotel reviews and applies a national standard (GB/T 14308–2023) for topic alignment, which may limit cross-regional generalizability. The reliance on full-review classification, rather than sentence-level sentiment separation, may overlook mixed-opinion nuances. Furthermore, the exclusion of reviews with model disagreement might introduce selection bias. Lastly, while ChatGPT and DeepSeek enhance topic validation, the lack of human adjudication may affect interpretive accuracy. Future research could adopt multilingual data sets, cross-standard mapping and hybrid annotation methods to improve adaptability and robustness.
This research pioneers the integration of deep semantic modeling (BERT and BERTopic) with standardized industry lexicons in the context of Chinese-language user reviews, offering a reproducible, interpretable and domain-aligned approach to analyzing hotel reviews. Beyond the luxury hospitality sector, the framework’s combination of deep learning classification, BERTopic clustering and semantic mapping to industry standards can be adapted to other service industries – such as healthcare, retail or transportation – where customer experience data is text-rich and domain-specific. The study introduces a dual-model cross-validation mechanism to ensure semantic rigor and presents a semantic mapping framework that bridges user sentiment data with operational evaluation systems, providing a scalable methodology for intelligent service optimization across diverse high-contact service environments.
