Entropy-aware filtering and hierarchical attention: A hybrid neural framework for fine-grained user satisfaction modeling.

With the rapid proliferation of internet-based applications, the volume of user-generated reviews has grown dramatically. Effectively extracting users' perspectives and emotional dispositions from this massive, heterogeneous text data poses significant challenges. This paper presents a novel multi-dimensional satisfaction analysis framework that integrates sentiment analysis with fuzzy decision-making. The sentiment analysis module comprises three key components: (i) an entropy-based mechanism for information usefulness prediction (IUP), (ii) hierarchical feature extraction using PERT-enhanced BERT encoders, and (iii) a sentiment value prediction model based on BERT-wwm-ext with a hierarchical attention-gated network. The fuzzy decision-making component leverages attribute weighting and an enhanced aggregation ranking strategy to compute user satisfaction. The proposed framework addresses limitations of traditional vote-based usefulness metrics and mitigates noise from redundant content. Extensive experiments on 18,559 user reviews from mobility travel service systems (MTSS) demonstrate that the model outperforms 12 baselines by 1.4%-5.6% across accuracy, F1-score, and AUC. This work provides a robust approach for fine-grained user emotion modeling and multi-attribute satisfaction evaluation in complex real-world scenarios.