<p>This paper discusses how to improve the long-sequence modeling and generalization ability across different music styles in the symbolic music generation task with limited data samples. First, we propose a neural network training approach based on meta-learning to enhance the generalization ability of models across different music styles in the symbolic music generation task with limited data samples. On this basis, we combine the meta-learning approach with Transformer-XL to improve the long sequence modeling capability of model, thereby optimizing the overall structure of composing music. Finally, the effectiveness of the proposed method is validated through subjective and objective evaluations. The experimental results indicate that: (1) within the same meta-learning framework, MAML-XL exhibits superior music sequence modeling and generation capabilities, significantly surpassing the baseline models in terms of scale consistency, pitch histogram entropy, pitch class histogram entropy, and groove consistency; (2) human ear evaluation results confirm the superiority of MAML-XL in the symbolic music generation task. The generated music performs well in terms of structure, rhythm, melody, audibility, and overall quality; (3) the ablation experiments demonstrate that meta-learning approach significantly enhances the quality of music generated by the model under the limited data condition.</p>