Audience Feedback Assessment System for International Communication of Dongguan Intangible Cultural Heritage Using Multimodal Sentiment Analysis CNN-BiLSTM Model

The digital dissemination of intangible cultural heritage (ICH) faces significant challenges in cross-cultural contexts, where diverse audiences exhibit varying sentiment patterns influenced by cultural backgrounds, language barriers, and content accessibility. Traditional sentiment analysis models, predominantly trained on Western datasets, fail to capture the nuanced emotional responses of global audiences toward culturally specific content, leading to biased assessments that disadvantage non-English speaking users. This paper presents a novel multimodal cross-cultural sentiment analysis framework specifically designed for ICH video promotion on social media platforms. We propose a CNN-BiLSTM fusion architecture integrated with a Cross-cultural Sentiment Calibration (CCSC) layer that combines textual comments, visual features, and cultural embeddings to achieve equitable and accurate sentiment classification across diverse cultural groups. The CCSC layer employs adversarial training to eliminate cultural bias while preserving sentiment-discriminative features, ensuring fair performance across language communities. We construct a comprehensive dataset comprising 4,643 comments from YouTube, Facebook, and Instagram on three representative Dongguan ICH items: Dragon Boat Racing, Incense Making, and Cantonese Opera. Experimental results demonstrate that our model achieves an F1-score of 87.3%, representing a 14.2 percentage point improvement over text-only baselines and a 4.3 percentage point gain compared to models without cultural calibration. The Cultural Fairness Index reaches 0.94, with performance standard deviation across language groups reduced from 5.8% to 1.1%—an 81% reduction in inequality. Comprehensive case studies reveal distinct cross-cultural reception patterns: action-based ICH (Dragon Boat Racing) achieves 92.3% positive sentiment with 89.1% acceptance rate, demonstrating strong universal appeal, while performance-based ICH (Cantonese Opera) shows only 43.7% acceptance despite 68.2% positive sentiment among engaged viewers, indicating significant language and cultural barriers. High-frequency term analysis identifies visual appeal, cultural authenticity, and skill demonstration as key drivers of positive sentiment, while language comprehension, pacing issues, and cultural context deficiency constitute primary engagement barriers. Our framework provides real-time audience feedback analysis with 200ms latency per comment, enabling ICH institutions to develop evidence-based promotion strategies that maximize cross-cultural acceptance. This research contributes a transferable methodology for cross-cultural sentiment analysis applicable to diverse cultural communication scenarios, advancing both the theoretical understanding of cultural bias in AI systems and the practical application of multimodal deep learning in digital humanities and cultural heritage preservation.