Multimodal Representation Learning for Winter Ice and Snow Wellness Tourism Experience Evaluation under Deep Learning

Tourism experience evaluation is often constrained by subjective assessment and insufficient integration of multimodal information. This study proposes a multimodal representation learning (MRL) framework for analyzing experiences in winter wellness tourism. The framework constructs hierarchical representations of textual, visual, and sequential information using Bidirectional Encoder Representations from Transformers (BERT), ResNet-50, and Bidirectional Long Short-Term Memory (BiLSTM) networks, respectively. Based on these representations, a cross-modal semantic alignment approach constrained by experience consistency is developed to reduce semantic discrepancies among different modalities. Furthermore, a dynamic multimodal fusion mechanism based on experience-state awareness is introduced to adaptively adjust the contribution of each modality according to changes in tourism scenarios, thereby enabling more effective characterization of dynamic tourism experience patterns. Experiments conducted on a real-world dataset comprising 50,000 reviews and 20,000 images show that the proposed model achieves an accuracy of 89.7% in sentiment analysis and 91.2% in satisfaction prediction, outperforming unimodal and late-fusion baselines. Ablation experiments further demonstrate the contributions of multimodal fusion and the attention mechanism to overall model performance. These findings suggest that the proposed framework provides a promising approach for multimodal experience analysis and intelligent tourism analytics.