The recommendation system that can assist rehabilitation doctors to provide objective and effective rehabilitation training programs for the huge patients with physical disability is a hot spot of current research. Traditional rehabilitation training program recommendation methods often ignore the contextual semantic information and potential semantic information in physical evaluation, resulting in inaccurate semantic expression and low accuracy of training program recommendation effect. Based on this, this paper proposes a deep learning network model (ALBERT-LDA model) based on ALBERT and LDA, and constructs a recommendation system for physical exercise rehabilitation training based on this model. The model is applied to construct a rehabilitation training program recommendation system. The recommendation system firstly uses the LDA topic model to obtain the document level topic information and the ALBERT model to obtain the word level semantic representation. Secondly, TextCNN is adopted to achieve feature extraction and dimension reduction of the word level semantic of the body evaluation text, and then hierarchical attention mechanism is used to reconstruct the dimensionality reduction features. Finally, the reconstructed features are fused with the subject features through certain fusion strategies to realize the recommendation of training schemes. A series of experimental results show that the limb movement rehabilitation training recommendation system constructed in this paper can not only obtain deeper semantic features through the fusion model, but also understand the context semantics more comprehensively. Therefore, it can describe complex semantic information in rehabilitation evaluation more accurately and improve the performance of multi-label classification. Thus, it can provide patients with more objective and effective rehabilitation training programs and improve the performance of medical recommendation system.