Abstract: To address the limitations of existing course evaluation models-specifically their insufficient sensitivity to cross-sentence context and inadequate extraction of semantic importance-this paper proposes a multi-granularity semantic analysis-based model that more accurately captures students' true intentions within textual feedback and supports downstream tasks such as sentiment classification and knowledge extraction. The model integrates pre-trained models and deep neural networks to extract word-level, sentence-level, and part-of-speech-level text feature vectors for semantic analysis and processing. Using course evaluation texts as an example, we conduct experiments and analyses. The model employs both precise and fuzzy matching evaluation methods and incorporates Dropout and the ReLU ( Rectified Linear Unit) activation function to enhance its generalization capability. In the experiments, we improved the model's classification performance by adopting various text preprocessing strategies, including stopwords removal and key term selection. The results indicate that the proposed model excels in sentiment analysis for course evaluations, achieving an accuracy of 92. 53% ,particularly when dealing with ambiguous sentiment boundaries. For course evaluations, the proposed semantic analysis model effectively captures detailed feedback from students, providing an efficient automated evaluation tool for the education sector and optimizing teaching quality.

Key words: