Abstract: The fluency and expressiveness of dance creation are often influenced by music rhythm and dance movement primitives. However, the existing methods only focus on a single influencing factor to generate dance movements, and it is difficult to realize multimodal information-driven dance movement generation. In order to solve the above problem, a dance generation method based on dance movement primitives and music semantics is proposed. First, a dance dataset is constructed. Unlike the mainstream dataset which only contains music and dance genre annotations, this dataset adds more fine-grained dance movement primitives on top of the existing annotations, which provides a data basis for dance generation driven by style, music and movement primitives together; Then, a diffusion model-based dance generation network is proposed, which is able to support the creative expression of a specific movement and reduce randomly generated movements through the introduction of the movement primitives feature extraction module. Finally, a set of optimized loss functions is designed to ensure the smoothness and physical realism of the dance through the loss of movement smoothness, the loss of movement continuity and the loss of physical reasonableness. Quantitative experiments on different datasets show that the proposed method outperforms comparative dance movement generation methods in evaluation metrics such as foot contact scores; qualitative experiments demonstrate the influence of factors such as dance primitives symbols or dance styles on the experimental results, which indicate that the proposed method is able to control the dance movements and their styles effectively when generating dance movements. The qualitative and quantitative evaluations show that the proposed method outperforms the existing methods in all evaluation criteria.