Emotion analysis and management improvement of hotel user online reviews based on BERT-RF and PLS-SEM integration

Online reviews have become a core factor affecting hotel booking decisions and brand reputation. Existing research mostly focuses on emotion analysis techniques, but lacks guidance on the implementation of management strategies, and traditional single text analysis has limitations. Therefore, to accurately locate hotel service issues, reduce the impact of negative reviews, enhance customer trust and platform ratings, an emotion analysis framework and real-time feedback mechanism that integrates multi-source text features are proposed. This study combines Bidirectional Encoder Representations from Transformers with Random Forest (BERT-RF) for emotion analysis, and uses structural equation modeling to model satisfaction and calculate satisfaction index. The emotion classification accuracy of the BERT-RF model reached 90%, with a positive review recall rate of 93% and an F1 score of 91%. The value perception had the strongest impact on satisfaction, with customer satisfaction directly driving loyalty and customer complaints negatively affecting value perception. The study accurately identifies negative reviews through multi-modal emotion analysis, and combines PLS-SEM quantitative satisfaction driving mechanism to provide data support for optimizing service resource allocation in hotels, promoting the transformation from “passive response to negative reviews” to “active prevention” management paradigm. Unlike conventional hybrid models that merely aggregate features, the proposed BERT-RF framework introduces a structured fusion mechanism. The deep context embedding of BERT is dynamically weighted and enhanced based on random forest statistics and domain-specific features, synergistically enhancing the robustness of sentiment classification, especially for fuzzy and domain-specific expressions in hotel reviews.