This paper proposes a multi-level discourse coherence evaluation framework aimed at addressing the complex challenges in assessing the coherence of Chinese essays. By integrating advanced deep learning technologies, including TextRCNN-LERT, UIE, and GLM4 models, we successfully constructed a comprehensive evaluation process from logical error detection to topic modeling and feedback generation. Experimental results show that this framework not only effectively identifies logical errors in essays, accurately extracts topic sentences and evaluates their logical relationships, but also generates specific and targeted feedback suggestions, significantly improving the accuracy and practicality of Chinese essay coherence evaluation.