Source-linked AI summary
Jiuge-Tuiqiao: An Interpretable Human-AI System for Classical Chinese Poetry Refinement
Yufeng Han, Lifan Deng, Cunliang Kong, Wenhao Li, Xin Cong, Yuzhuo Bai, Kangyang Luo, Maosong Sun
TL;DR
One-shot poetry generation can weaken creator agency and offers opaque refinement. Jiuge-Tuiqiao instead combines user controls, classical evidence, AI assistance, and prosody feedback; preliminary evaluation and user studies report improved creative agency and engagement.
Problem
One-shot AI poetry systems limit local iterative refinement, obscure recommendation rationale, and reduce users’ creative agency.
Method
Jiuge-Tuiqiao uses a triadic human-AI workflow with character or line locking, real-time prosody feedback, and evidence-grounded suggestions.
Results
User studies report significantly enhanced creative agency and engagement, while the BOTH ablation achieves 60% Hit@5 and 56% Hit@1.
Takeaways & Limitations
The system positions AI as a background assistant that supports human-led poetic refinement rather than replacing creative agency.
Takeaways & Limitations
The knowledge base is skewed toward Tang-Song traditions and elite male authorship, under-representing women poets, border literati, and minority-language contributors.
Abstract
from arXiv · showhide
Classical Chinese poetry composition has long valued Tuiqiao, the iterative refinement of words, imagery, and prosody. However, many current AI poetry systems follow a one-shot generation paradigm, which reduces users to prompt providers and weakens their creative agency. We present Jiuge-Tuiqiao, an interactive human-AI collaborative system for classical Chinese poetry composition. The system is designed around a triadic model: user-driven control, ancient-guided evidence, and AI-assisted generation. Users can lock characters or lines, receive real-time prosody feedback, and obtain interpretable refinement suggestions grounded in high-frequency collocations, PPL-ranked classical lines, and structured knowledge extracted from classical encyclopedias. This design turns AI from an autonomous generator into a background assistant that supports the user's own process of poetic refinement. Preliminary experiments and user feedback suggest that Jiuge-Tuiqiao improves controllability, interpretability, and user engagement in classical poetry composition.
1 Introduction
Jiuge-Tuiqiao addresses one-shot poetry systems by restoring iterative Tuiqiao, creator agency, and interpretable classical guidance. Its triadic workflow combines user control, ancient knowledge, AI assistance, and real-time prosody feedback.
- Tuiqiao refines poetic words, imagery, and prosody while preserving the creator’s linguistic competence, cultural literacy, and stylistic voice.The Jia Dao example illustrates balancing linguistic accuracy, emotional nuance, and formal integrity during refinement.
- Current one-click systems lack precise local intervention, making iterative refinement—the essence of Tuiqiao—impossible.
- Opaque recommendations provide no literary or linguistic justification, preventing users from understanding or verifying candidate lines against classical precedents.
- One-shot interaction reduces users to instruction-givers and passive receivers, weakening authorship and creative composition.
- Jiuge-Tuiqiao models collaboration among the user, ancient knowledge, and AI, with users locking text and receiving constrained revision suggestions.
- Real-time color-coded feedback on tonal patterns and rhyme helps users attend to prosody and semantic expression during refinement.
- User studies report significantly enhanced creative agency and engagement in a fully implemented WeChat mini-program workflow.
2 Related Work
Related work spans automated poetry generation and interactive writing systems, while Jiuge-Tuiqiao emphasizes interpretable, evidence-grounded refinement in classical Chinese poetry.
- Automated Chinese poetry generation has progressed from rule-based methods to neural architectures optimizing coherence and style.
- Recent systems add structural, imagery, character, and domain-specific adaptation controls through GPT-2, PoemBERT, CharPoet, and LoRA methods.
- Interactive platforms support iterative draft revision but their recommendations remain model-driven and lack literary justification.
- Jiuge-Tuiqiao grounds refinement in ancient encyclopedias and historical collocations, extending feedback-driven human-AI intervention to constrained classical poetry.
3 System Architecture
Jiuge-Tuiqiao formalizes poetry refinement as a constrained, iterative human-AI process implemented through a three-layer architecture and a finite-state interaction loop.
- Problem Formulation: The system represents a poem as a partially observable character matrix whose dimensions depend on poetic form and whose entries may be placeholders.
- Problem Formulation: Users define constraints comprising keywords, artistic description, rhyme category, optional tonal pattern, and a verse-derived tonal matrix.
- Problem Formulation: A lock matrix freezes selected characters while the objective optimizes an implicit LLM score under user constraints and prosodic requirements.
- Three-Layer Architecture: The three-layer architecture separates user interaction, classical knowledge, and constraint inference to maintain user control and generate compliant candidates.
- Iterative Tuiqiao Process: The FSM cycles through initialization, AI generation, user review and locking, and repeated refinement until the user is satisfied.
- Iterative Tuiqiao Process: AI generates interpretable options while historical evidence supports them, leaving acceptance, rejection, editing, and creative agency with the human user.
4 Methods
The method retrieves interpretable refinement evidence from collocations, famous lines, and ancient encyclopedias, then presents traceable candidates for user-directed editing.
- Evidence Sources: Jiuge-Tuiqiao retrieves frequent co-occurrences, famous couplet matches, and ancient encyclopedia clues for each refinement position.
- Frequent Co-occurrence Retrieval: Collocations are extracted from 320,000 Tang-Song poetry lines by counting word and ordered-pair frequencies after corpus preprocessing.
- Frequent Co-occurrence Retrieval: Pairs with f(w_i, w_j) ≥3 and T > 1.96 yield 28,739 statistically filtered high-frequency collocations.
- Frequent Co-occurrence Retrieval: Candidates are ranked by their highest t-value with context words and returned alongside traceable collocation evidence.
- Famous Couplet Matching via Perplexity: The system uses PPL as a proxy for canonical familiarity, assuming lower-PPL lines are more likely to reflect canonical historical expressions.
- Famous Couplet Matching via Perplexity: It computes PPL with Qwen3-32B over a collection built from more than 1.8M historical poems, then ranks candidates within length bins.
- Ancient Encyclopedia Guidance: Structured data from 11 ancient encyclopedias forms imagery, antithesis, and rhyme databases queried using keywords, current words, and target rhyme.
- Evidence Presentation: Categorized cards expose evidence from the knowledge sources so users can adopt candidates or manually edit the poem.
5 User Interface
Jiuge-Tuiqiao combines direct user control with real-time prosody feedback and traceable literary evidence. Its interface supports iterative refinement by letting users lock text, inspect suggestions, and progressively revise candidates.
- Real-Time Feedback: The Character Grid Editor provides color-coded feedback for valid compliance, acceptable variants, rhyme mismatches, tonal violations, and user-locked tokens.Explicit justifications accompany violations to support correction during composition.
- User Control: Users can lock individual characters or entire lines, preserving selected text while restricting AI generation to unlocked content.Character-level locks freeze cells, while line-level locks fix a complete row and constrain generation to remaining lines.
- Interpretable Evidence: The Interpretable Tuiqiao Panel presents co-occurrence candidates, ranked famous lines, encyclopedia clues, and cached LLM rationales in parallel.These channels provide traceable literary evidence and support deeper historical inspection through Sou-Yun links.
- Iterative Workflow: Users browse evidence cards, adopt candidates, manually edit text, and retain final decision authority throughout refinement.The workflow cycles through initialization, candidate generation, locking, and progressive revision with real-time validation.
6 Evaluation
The evaluation examines knowledge-source contributions, prosodic correctness and diversity, and user experience. Results show strong local prosody and diversity, weaker poem-level adherence for regulated verse, and positive user ratings.
- 6.1 Ablation Study: The controlled cloze-filling ablation compares BASE, +FL, +CO, and BOTH using Hit@1 and Hit@5 on 50 poem lines.Each instance masks a contiguous two-character content word for Qwen-turbo prediction.
- 6.1 Ablation Study: 60% is the highest Hit@1 for +FL, while BOTH reaches 60% Hit@5 and 56% Hit@1.Famous lines preserve classical diction, whereas co-occurrences expand lexical search and stylistic diversity.
- 6.2 Automatic Prosody Evaluation: Perfect Format and Rhyme accuracy coexist with L-Pattern ≥91.25% across generated poem forms.These metrics assess structural length, line-level tonal accuracy, and Pingshui Yun rhyme consistency.
- 6.2 Automatic Prosody Evaluation: P-Pattern adherence falls to 15%–25% for lvshi versus 50%–65% for jueju, reflecting compounded eight-line nian-dui requirements.The result identifies global regulated-verse constraints as a bottleneck.
- 6.3 Human Evaluation: Prosody guidance scores 8.20/9, overall satisfaction 8.08/9, agency 7.56/9, and trustworthiness and evidence utility 7.08/9.Ten participants completed a 9-point Likert questionnaire after composing at least one poem.
7 Conclusion and Future Work
Jiuge-Tuiqiao recenters classical poetry composition on human agency through a triadic, transparent collaboration among users, ancient knowledge, and AI. The authors report high prosodic compliance and stronger user trust, while identifying personalization, broader forms, larger studies, and preference alignment as future directions.
- The system implements a triadic design integrating hierarchical locking, real-time prosody validation, and multi-channel traceable recommendations.
- Evaluation and user studies report high prosodic compliance and significantly greater user trust than black-box alternatives.
- Future work targets personalized recommendations, irregular poetic forms, larger and more diverse user studies, and preference-aligned LLM fine-tuning.
Ethics and Impact
Jiuge-Tuiqiao preserves human initiative while embedding classical scholarship into an accessible poetry-composition workflow. The paper discusses disclosure and authorship, cultural accessibility, historical-corpus bias, and possible dependence on refinement assistance.
- The system leaves creative initiative with the poet, requiring user acceptance before inserting text and providing an export disclosure template for AI-assisted poems.
- Embedding Pingshui Yun, encyclopedias, and famous-line corpora may lower barriers for second-language learners and heritage communities outside China.
- Historical knowledge sources skew toward Tang and Song materials and elite male authorship, under-representing women poets, border literati, and minority-language contributors.
- Persistent refinement assistance may discourage independent prosodic intuition, while explanatory feedback and visible justifications are intended to encourage learning.
A Rhyme Categories and First-line Tonal Patterns
The appendix defines Pingshui Yun rhyme categories and represents seven-character jueju tonal constraints as matrices over level, oblique, and flexible positions. It consolidates four globally valid start-and-end tonal patterns.
- Rhyme categories: The Pingshui Yun system supports 30 Level Tone categories, divided into Upper Level Tones and Lower Level Tones.
- Tonal pattern matrices: Tonal constraints use matrices T ∈ {P, Z, A}^n×m, where P denotes level, Z oblique, and A flexible tone positions.
- First-line tonal patterns: The four valid seven-character jueju patterns combine level or oblique starts with level or oblique ends.
- First-line tonal patterns: The listed matrices specify position-level P, Z, and A assignments for level-start/level-end and level-start/oblique-end patterns.
- First-line tonal patterns: The appendix also lists the oblique-start/level-end and oblique-start/oblique-end pattern labels with their corresponding tonal sequences.
B Poem Segmentation Algorithm
The poem segmentation algorithm generates candidate paths using dictionary scores and metrical rules, then reranks the top candidates with an LLM under explicit grammatical guidelines. Its scoring combines lexical evidence, literary allusion, word length, vocabulary status, and metrical rhythm.
- Two-stage segmentation: The two-stage G-Seg strategy enumerates top-K segmentation paths with DFS before LLM reranking selects the grammatically best candidate.
- Dictionary scoring: The path score combines dictionary evidence with metrical bonuses, incorporating lexical features such as frequency, mutual information, dictionary counts, and allusions.
- Dictionary scoring: Length preferences reward two- and three-character words, penalize words of length four or more, and assign OOV tokens a base score of −1.0 plus a −0.1 penalty.
- Metrical constraints: Metrical bonuses model 2 + 3 rhythms for five-character lines and 2 + 2 + 3 rhythms for seven-character lines.
- LLM reranking: LLM reranking keeps functional words separate, merges tightly coupled entities, and prioritizes contextual flexibility over blindly selecting longer tokens.
C Ancient Encyclopedias Database
Jiuge-Tuiqiao builds a structured classical-knowledge database from 11 sourcebooks and uses it to provide traceable, interpretable refinement support. The database organizes imagery, antithesis, and rhyme information while exposing source evidence through interactive cards.
- Sourcebook Construction: The system constructs its knowledge base from over ten classical Chinese encyclopedias and rhyme dictionaries spanning the Tang to Qing dynasties.The utilized resources include 11 classical sourcebooks, including Chuxueji, Taiping Yulan, and Peiwen Yunfu.
- Data Processing: 503,908 structured records are consolidated into three SQLite index tables that provide explicit etymological anchors.The records are produced by parsing, cleaning, and rule-filtering texts from Siku Quanshu editions and verified open-source assets.
- Knowledge Organization: The image word database links thematic keywords to historically favored imagery tokens with definitions and cited literary contexts.It aggregates six sourcebooks and supports conceptual justification during theme refinement.
- Knowledge Organization: The antithesis word database extracts original parallel couplets and recommends position-specific classical pairings with source texts as references.It combines five phonological books and calibrates recommendations against corresponding characters in parallel lines.
- Knowledge Organization: The rhyme word database extracts Pingshui Yun category characters and associated phrases to supply rhyme-compliant generation candidates.These verified knowledge streams are injected as structural constraints and rendered through interactive cards for knowledge tracing.