<p>Recognition and understanding figure of speech is an important task. With the rapid development of large language models, substantial progress has been made in the recognition and understanding of rhetorical devices. However, existing methods largely rely on large-scale annotated data and shallow linguistic features, leading to limited generalization in low-resource settings; moreover, they remain insufficient in modeling the deep semantic interactions underlying rhetorical structures. This paper proposes a multidimensional character-level fusion approach for Chinese rhetorical device recognition. By systematically designing and integrating character-level elements derived from the composition and application of rhetorical devices, the method improves the model’s ability to recognize Chinese rhetorical expressions. The impact of different character-level fusion strategies on rhetorical recognition is also investigated. In a dataset of K12 textbooks covering eight commonly used speech figures, the proposed method has demonstrated strong capability in distinguishing multiple figures of speech, achieving an overall recognition rate of around 70% for common figures, a 4% improvement compared to baseline models. In addition, data indicate that character-level features, such as semantics, Pinyin, and character components play prominent roles in the recognition task. In the public dataset CERRE, the method proposed in this paper outperformed the 32B-scale large-language model. Ablation studies and analyses illustrate the influence of different character-level combinations on model performance, validating the effectiveness of the method while providing empirical insights for other Chinese understanding tasks such as linguistic interest mining, intelligent education, and literary research.</p>