Generation of stylized Chinese characters based on the structure and semantics of Chinese characters (Undergraduate Thesis)

Examples of novelly generated Chinese characters Left: Initial Combined Character created by the Dynamic Word-Combination Algorithm Right: Optimized Combined Character provided by the Character-Composition Optimization Network

Under the supervision of Professor Shibo He (Control Science, Zhejiang University) and Professor Kejun Zhang, I completed my undergraduate thesis Generation of Stylized Chinese Characters Based on Structural Semantics in the field of Chinese character generation by integrating Chinese character structure theory with artificial intelligence techniques. This work has also been written as a granted Chinese patent: Multi-style dynamic word combination method based on generative adversarial network (基于生成对抗网络的多风格动态组字方法), in which I am the first student author.

My major contributions include:

1. Proposed a novel method for automatic Chinese character generation based on Ideographic Description Sequences (IDS)—a semantic notation describing the structure and components of Chinese characters
2. Developed a two-stage generation pipeline:
   1) A dynamic word-combination algorithm that analyzes IDS and constructs initial character layouts
   2) A character-composition optimization network (GAN architecture )that refines the layout using deep learning
3. Trained the optimization network with a customized dataset
4. Demonstrated improvements of 24–48% on key metrics such as NRMSE and PSNR
5. Validated the method’s ability to generate both rare and artistic “novel characters”, such as ligatures and creative compositions