In the open composition review scenario for primary and secondary school students, there are no clear criteria and less labeled data that correlates the reviews with the grade level and the type of composition. To help students understand the reviews for improving the quality of composition, it is necessary to show the relevance between the reviews and the compositions content. In this paper, we propose the meta-instance incorporated composition review generation method, which learns from the similar compositions for reviewing new compositions. We employ a composition segmentation method to construct a meta-instance dataset that contains composition fragments and their corresponding reviews. To eliminate the issue of hallucination caused by the meta-instance, the cross-content detection method and the masking mechanism are designed. We also design the retriever model to find the relevant meta-instances with a new composition, where the text semantic encoder is trained by contrastive learning. Then the review for a new composition is generated by combining the content of the new composition with the meta-instances’ reviews. Experiments were conducted on a real-world composition review dataset, and the results demonstrated that our method outperformed the existing approaches. We also compare the diversity of reviews generated by different models.