The current English composition automatic correction system rarely involves coherence quality analysis of compositions. Therefore, a semantic coherence diagnosis model for English texts was constructed based on sentence semantic maps, and its effectiveness was verified through experiments. Experimental results show that sub Graphical model 10, 12 and 13 exceed 200 on the first two test texts and 1 on the last two test texts. However, the differences between these three subgraphs on different test texts are not significant, with differences below 30 and 0.3. In addition, when extracting incoherent sentences, the F1 value reaches the optimal value at a threshold of 0.34, which is 87.54%. When the threshold is fixed at 0.34, the accuracy of extracting non coherent sentences also increases with the number of articles, reaching a maximum of 88.43%. At the same time, there was no significant difference in accuracy, recall, and F1 values among different English composition numbers, maintaining between 83% and 89%. The Pearson coefficient calculated in the actual comparison with the teacher's manual composition score is 0.6025, indicating a strong correlation between the two, indicating that the diagnostic results are reliable. Overall, the diagnostic model constructed in the study has strong accuracy and effectiveness, and is practical in the diagnosis of semantic coherence in actual English texts.