A noise-robust unsupervised human skeleton representation learning method based on multimodal large models was proposed to address the challenge that existing methods struggled to capture subtle differences in actions (e.g., paper cutting and looking at a phone). The action description texts generated by multimodal large models were incorporated to provide key action information for training. A joint noise estimation approach based on outlier sample detection and information entropy was developed to provide high-quality auxiliary textual signals for skeleton representation learning and improve the noise robustness of multimodal large model. The experimental results on three public datasets showed that text-assisted learning enabled the model to significantly outperform the existing methods in downstream tasks of action recognition and retrieval, while demonstrating strong knowledge transfer capability of the model, particularly in low-resource scenarios. The method has good generalization ability and is suitable for most existing unsupervised skeleton representation learning frameworks, while its joint noise estimation method is also effective in cross-modal retrieval tasks.