Abstract: For current chapter graded consistency metric model only considers the full text consistency of the tested composition, and cannot capture the implicit semantic characteristics of the text language block and the consistency between them. In view of the above problems, a composition text consistency evaluation model was proposed for general compositions. Refering to the thoughts of the twin neural network, the model was arranged to extract the character, image characteristics, and storyline characteristics of the core character simultaneously in the composition and to perform similarity metrics, so as to obtain the central idea of the text and the matching score of the text consistency. A subject model Biterm-LDA（Latent Dirichlet Allocation）was used to extract the subject character of composition to avoid the dependence of the artificial labeling. The results show that the proposed model score is consistent with the results of artificial labeling, and is superior to ordinary neural network models.