Jiuge-Tuiqiao: An Interpretable Human-AI System for Classical Chinese Poetry Refinement
This paper presents Jiuge-Tuiqiao, an interpretable human-AI collaborative system that shifts classical Chinese poetry composition from one-shot generation to an interactive refinement process by empowering users with real-time prosody feedback, ancient-guided evidence, and controllable AI suggestions.
Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
For centuries, the creation of classical Chinese poetry has been less about a sudden flash of inspiration and more about a deliberate, painstaking process of refinement. This practice, known as Tuiqiao, involves a poet carefully weighing individual words, testing how they sound against strict rhythmic rules, and polishing imagery until every line feels inevitable. It is a craft where the writer acts as both the architect and the editor, constantly revising a draft to balance meaning, sound, and tradition. In the modern era, artificial intelligence has become remarkably skilled at writing poetry, often producing fluent verses in a single attempt. However, this speed comes with a cost: the human writer is reduced to a passive observer, handing a prompt to a machine and receiving a finished product without the ability to steer the creative process or understand why the machine chose specific words. This disconnect leaves the writer without the agency that defines the traditional art form.
Researchers at Tsinghua University have developed a new system called Jiuge-Tuiqiao to bridge this gap, creating a tool that returns the creative initiative to the human poet. Rather than generating a complete poem in one go, the system functions as a collaborative partner that supports the iterative process of refinement. It operates on a simple but profound idea: the human remains the director, while the artificial intelligence acts as a knowledgeable assistant that offers suggestions grounded in historical evidence. The system allows a user to lock in specific characters or lines they are satisfied with, effectively telling the computer, "Keep this part, and help me figure out the rest." As the user works, the system provides real-time feedback on the poem's rhythm and rhyme, using color codes to highlight errors or confirm compliance with ancient rules.
What makes this system distinct is how it justifies its suggestions. Instead of offering a list of random alternatives, Jiuge-Tuiqiao pulls evidence from three specific sources to explain why a word might fit. First, it looks at high-frequency word pairings found in hundreds of thousands of Tang and Song dynasty poems, showing the user which words historically appear together. Second, it retrieves famous lines from history that contain similar phrases, allowing the user to see how masters of the past handled comparable situations. Third, it draws from ancient encyclopedias to offer clues about imagery, antithesis, and rhyme categories. When a user clicks on a blank space in their poem, they do not just see a list of options; they see a panel of traceable evidence, including the original source of a famous line or the statistical likelihood of a word pairing. This transparency turns the interaction into a learning experience, where the user can verify the machine's logic against the weight of literary tradition.
The researchers tested this approach to see if it truly improved the creative experience. In one set of experiments, they asked the system to fill in missing words in existing poems under different conditions. When the system relied only on the immediate context, it struggled to find the correct words. However, when it was allowed to use the historical evidence—specifically the famous lines and word pairings—its ability to suggest the right words improved significantly. The combination of statistical patterns and canonical examples proved most effective, suggesting that the system successfully mimics the way a human poet might consult a mental library of references. Furthermore, the system demonstrated a high degree of accuracy in adhering to the strict tonal patterns and rhyme schemes required by classical forms, achieving near-perfect scores on structural rules while still maintaining a wide variety of vocabulary.
Human participants, including poetry enthusiasts and researchers, tested the system in a practical setting. They reported that the real-time feedback on rhythm and rhyme was particularly useful, helping them correct mistakes instantly. More importantly, the users felt a strong sense of control over the final work. They trusted the suggestions more when they could see the historical evidence behind them, preferring the system's ability to cite ancient precedents over simple statistical guesses. The study suggests that by making the AI's reasoning visible and keeping the human in the driver's seat, the system successfully preserves the spirit of Tuiqiao. It does not replace the poet but rather extends their capacity to refine their work, turning the act of writing poetry back into a dialogue between the creator, the machine, and the centuries of literature that came before.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.