BERT-TextCNN Fusion Model for Folk Culture Text Sentiment Analysis and Practice Classroom Feedback Optimization in Greater Bay Area

Abstract

Folk cultural activities in the Greater Bay Area generate a large amount of unstructured text in social media and classroom practice. The emotional expressions in these texts often possess colloquial, metaphorical, and dialectal characteristics, making traditional sentiment analysis insufficient to support the refined needs of "cultural identity cultivation—classroom iterative optimization." Addressing the shortcomings of existing methods in multidimensional attitude characterization, unstable cross-language generalization, and weak interpretability of teaching decisions, this paper proposes a BERT–TextCNN fusion framework for dual-domain texts. This framework captures context-dependent information using pre-trained semantic representations and enhances the identification of emotional trigger segments and fixed collocation structures by combining convolutional local pattern extraction. Simultaneously, multi-task learning achieves joint modeling of three-dimensional emotions: interest, participation willingness, and cultural identity. Under dual-domain evaluation and cross-domain transfer settings, the proposed method outperforms strong baselines and single pre-trained models in metrics such as macro-average F1, with less cross-domain degradation. Furthermore, mapping multidimensional emotional outputs to optimize classroom content and interaction strategies significantly improves indicators related to cultural identity and participation. This paper contributes by constructing a multi-dimensional sentiment modeling paradigm for folk culture education scenarios, verifying the robustness of the fusion structure in complex contextual texts, and providing a reusable closed-loop path for "sentiment insight-driven classroom feedback optimization," thus offering methodological support for data-driven improvement of traditional culture education.