<p>The evaluation of dance expressiveness has traditionally relied on expert judgment, which is inherently subjective and often inconsistent. Existing artificial intelligence (AI)-based approaches are further limited by their “black-box” nature and insufficient depth in multimodal fusion. To address these limitations, this study proposes an intelligent evaluation framework for dance expressiveness that integrates multimodal data with explainable artificial intelligence. Specifically, the Dance Performance Multi-modal Dataset (DPMM-Dataset) was constructed, comprising 480 video clips from 120 dancers across four major dance genres, incorporating visual, audio, and skeletal modalities. The proposed model adopts a hierarchical dual-path attention mechanism, enabling fine-grained decomposition of technical components (e.g., limb extension) while facilitating effective cross-modal feature integration. To enhance interpretability, Shapley Additive Explanations (SHAP) are employed to quantify the contribution of individual features, thereby improving transparency in the decision-making process. Experimental results indicate that the proposed model significantly outperforms baseline methods in terms of Mean Squared Error (MSE) and the coefficient of determination (R²). Furthermore, the model demonstrates strong generalization across all four dance genres, achieving up to 92.1% accuracy in technical dimension evaluation. Overall, this study provides a practical and interpretable framework for the digital and standardized assessment of dance performance, offering both theoretical insights and practical value for dance education and competition evaluation.</p>