Research on Experience Evaluation and Optimization of Rural Festival Participants Using BERT Multimodal Sentiment Analysis

Abstract

To solve the problems of separating emotional perception from textual and image commentary in rural festival activities and identifying multi-dimensional experience pain points, particularly those ‘implicit’ contradictions where positive textual sentiment masks negative visual cues (e.g., a visitor praising the scenery while the image reveals severe over crowding), remains challenging. This paper proposes a multimodal sentiment analysis framework that combines BERT semantic encoding and cross-modal collaborative attention. First, the textual and image data of typical rural festivals in the Yangtze River Delta region are collected and processed to align the two modalities. Second, BERT is finetuned to extract textual semantics, and ResNet-50 is employed to encode visual atmosphere. Finally, a collaborative attention mechanism is introduced to integrate text and image data, allowing the model to simultaneously process cross-modal interactions and pinpoint the specific visual sources of participant sentiment. The experimental results show that the fusion model obtains a macro-average F1 score of 91.5%, which is an improvement of 9.1% compared with the singlemodal model. Furthermore, the model can also locate the core pain points of crowdedness with the attention weight of 0.78, which provides an interpretable basis for optimizing rural festivals.