A Time-Aware Sliding Window-Based Hotel Recommendation Framework Using Multi-Stage BERT-MRC

User-generated reviews offer very valuable insight into service quality, but these reviews are time sensitive; that is, old reviews are less important than new ones. In this work, we propose a time-aware hotel recommendation framework that combines fine-grained aspect-based sentiment analysis with a sliding window mechanism to capture evolving customer feedback. We define a fixed set of hotel-related aspects and adopt a three-stage BERT-based Machine Reading Comprehension (MRC) pipeline. The first model extracts aspect terms from reviews, the second identifies corresponding opinion terms, and the third predicts the sentiment polarity of each aspect based on its associated opinions. The extracted aspect terms are then modeled on the predefined aspects using semantic similarity. Hotel profiles are constructed by aggregating aspect-wise sentiment scores over sliding time windows, with more recent windows assigned higher weights. Reviews within each window contribute equally, and the total weight across all windows sums up to one. User preferences are represented in the same aspect space, and hotel recommendations are generated based on cosine similarity between user preferences and hotel profiles. Experimental results show that the proposed system effectively captures both the fine-grained opinions and their temporal relevance, providing a solid foundation for dynamic and personalized hotel recommendations.