Introduction
Chinese character riddles, known as zi mi, represent a sophisticated cultural tradition where the solution is always a single Chinese character. Unlike Western riddles that primarily leverage semantic puns, zi mi are structural puzzles. They describe the composition of a character by decomposing it into radicals or using poetic metaphors to hint at its constituent parts. For instance, a riddle might describe a "woman" and a "horse" meeting to form the character for "mother" (妈).
The computational challenge lies in two areas: solving existing riddles by mapping poetic phrases to structural components, and generating new, elegant riddles that maintain traditional literary standards.
System Architecture
The researchers developed a unified statistical framework to address both solving and generation. The system operates in three primary phases: offline learning, riddle solving, and riddle generation.
Offline Learning
The foundation of the system is a corpus of 77,308 riddle-solution pairs mined from the web. Using statistical alignment techniques—analogous to those in machine translation—the system learns associations between phrases and radicals. For example, it identifies that "west lake" often corresponds to the "water" radical (氵). Additionally, 193 structural rules (e.g., "HalfRemove") were extracted to describe character manipulations.
Riddle Solving
Solving is treated as a parsing problem. The system employs a dynamic programming algorithm based on the Cocke-Younger-Kasami (CYK) approach to explore possible combinations of phrases into radicals and radicals into characters. To manage the large search space, a Ranking SVM scores candidates based on features such as radical match frequency and the commonality of the resulting character.
Riddle Generation
Two strategies were implemented for generation:
- Template-based: Decomposes a target character into radicals and retrieves corresponding metaphors from the learned alignment table, connecting them with standard conjunctions.
- Replacement-based: Adapts existing human-written riddles by swapping components for a new target character, ensuring higher fluency and naturalness.
Experimental Results
The system was evaluated against traditional translation baselines and human performance.
- Solving Performance: On a test set of 2,000 riddles, the framework achieved a Top-1 accuracy (Acc@1) of 28.8%, significantly outperforming the GIZA++ baseline of 10.3%. The Top-10 accuracy reached 32.1%.
- Generation Quality: Human judges rated generated riddles on a 5-point scale. The replacement-based method achieved a score of 4.14, nearly matching the 4.38 score of human-authored riddles.
Conclusion
This research presents the first comprehensive statistical approach to the unique domain of Chinese character riddles. By focusing on structural assembly and metaphorical mapping rather than pure semantics, the system demonstrates that AI can both solve complex linguistic puzzles and generate culturally resonant creative content. Future work may leverage deep learning architectures to further enhance the poetic elegance and semantic depth of generated riddles.
Summary updated