基于预训练语言模型的汉语字形谜解谜模型

Summary

The paper addresses automatic solving of Chinese character-shape riddles (字形谜), in which the clue describes a character's components and their spatial arrangement. The authors note that general LLMs handle these poorly; GPT-4, for example, gave 玥 instead of 金 for a "人 above 玉" clue. To fill the data gap, they built a character resource from two sources. Basic attributes such as strokes, pinyin and radical for about 3,500 common characters came from the mapull/chinese-dictionary project, and decompositions came from kfcd/chaizi. Riddles were then generated automatically by filling structure-specific templates with each character's components.

The solver runs as a pipeline with three steps:

• a BGE encoder with a small feed-forward classifier predicts the character's structure type;
• a fine-tuned BART model extracts the key components;
• a lookup dictionary maps the structure and components to the answer character.

On a test set built from templates never seen in training, the system reaches 96.90% overall accuracy. This shows that fine-tuned pre-trained models can capture sub-character structure, with possible uses in language teaching and cultural applications.

Key points

• Structure types: The dataset covers five layouts (top–bottom, left–right, top–middle–bottom, left–middle–right, full enclosure).
• Split: Data is divided 7:3 into 1,052 training and 452 test riddles. The heavy skew toward left–right and top–bottom characters is offset with template-generated pseudo-data and oversampling.
• Accuracy: Structure prediction reaches 98.92%, keyword extraction 97.30%, and the combined answer 96.90%. Character accuracy equals the combined figure.
• Error sources: Errors stem from rare components such as 丩, from function words being skipped (而 misread), and from confusing top–middle–bottom with top–bottom clues.
• Code: The code is released openly on GitHub (repository "zimi").