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

cs.AIAug 24, 2026

Jiuge-Tuiqiao: An Interpretable Human-AI System for Classical Chinese Poetry Refinement

Authors: Yufeng Han, Lifan Deng, Cunliang Kong, Wenhao Li, Xin Cong, Yuzhuo Bai, Kangyang Luo, Maosong Sun

Organizations: Department of Computer Science and Technology, Tsinghua University, Beijing · Beijing National Research Center For Information Science And Technology · Institute for Artificial Intelligence, Tsinghua University, Beijing · Rixin College, Tsinghua University ·

Abstract

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.

Figures & tables

Panel of Figure 1 — Figure 1: The collaborative workflow of Jiuge-Tuiqiao . It shows the relationship where the User maintains creative subjectivity, Ancient Knowledge provides knowledge guidance, and AI offers generative assistance.

Panel of Figure 2 — Figure 2: System architecture and the triadic collaborative iterative Tuiqiao process. The core engine coordinates the Ancient Heritage, User Subjectivity, and AI Assistant layers through continuous feedback and evidence retrieval.

Figure 3: Overview of system states: (a) metadata constraint initialization panel; (b–e) real-time feedback in the Character Grid Editor showing (b) valid compliance (green) and acceptable variants (yellow), (c) rhyme mismatch (purple), (d) tonal violation (red), and (e) user-locked tokens (gold fill).

Figure 4: Multi-channel evidence in the Interpretable Tuiqiao Panel: (a–b) statistical high-frequency co-occurrence candidates, (c) ranked famous-line samples with metadata, and (d) textual clues retrieved from ancient encyclopedias.

Condition	Hit@1	Hit@5
BASE	12%	12%
+FL	60%	60%
+CO	36%	42%
BOTH	56%	60%

Table 1: Ablation results on cloze-filling. Combining both sources ( BOTH ) matches the highest Hit@5 (60%) while yielding a slightly lower Hit@1 (56%) than +FL alone (60%).

Form	Format	L-Pattern	P-Pattern	Rhyme
5-char jueju	100%	96.25%	65%	100%
7-char jueju	100%	91.25%	50%	100%
5-char lvshi	100%	93.12%	25%	100%
7-char lvshi	100%	93.12%	15%	100%

Table 2: Automatic prosody accuracy by verse form.

	1-gram	2-gram	4-gram
Distinct	0.9804	0.9995	1.0000
Self-BLEU	0.8983	0.5717	0.1855

Table 3: Output diversity metrics across generations.

Figure 5: Human evaluation ratings (9-point Likert scale) showing (a) four evaluation dimensions and (b) perceived utility of three knowledge sources.

Appendix figures & tables2 assets

Supplementary material from the paper’s appendix.

Appendix

Line Type	Boundary Rules	Length Rules
5-char	Pos 2: $+0.8$	2-char: $+0.1$
5-char	Pos 3: $-0.3$	1-char: $-0.05$
7-char	Pos 2, 4: $+0.5$	2-char: $+0.1$
	Pos 3, 5: $-0.3$	1-char: $-0.05$
		$\geq 4$ -char: $-0.1$
General	-	2-char: $+0.05$

Appendix

Table 4: Metrical rhythm reward and penalty rules.

Book	Description	Content Example (Extracted Data)
《北堂书钞》 Beitang Shuchao	Compiled by Yu Shinan (Tang). An early extant encyclopedia focusing on politics and rituals, preserving numerous pre-Sui texts.	Keyword: 功业 (Achievements) Volume: 4 Section: 帝王 (Emperors) Content: “四本具即帝初立，举而措之事业，功业赫赫，功盛德厚，功侔太古…”
《白孔六帖》 Baikong Liutie	Compiled by Bai Juyi (Tang) and Kong Chuan (Song). Collects phrases and sentences as source materials for poetry and essays.	Keyword: 天 (Heaven) Volume: 1 Content: “(白)高明柔克(髙明天也柔克寒暑不干), 阴骘下民(言天黙定下民之命)…”
《太平御览》 Taiping Yulan	A massive state-sponsored encyclopedia from the early Song Dynasty. It is highly comprehensive, systematically categorized by heaven, earth, human, objects, etc.	Keyword: 太初 (Vital Energy) Volume: 1 Section: 天部一 (Heaven) Content: “《易乾鉴度》曰：太初者，气之始也。《帝王世纪》曰：元气始荫，谓之太初…”
《艺文类聚》 Yiwen Leiju	Compiled by Ouyang Xun (Tang). A state-sponsored encyclopedia intertwining historical facts and literature (poems and prose) under a clear Confucian orthodox viewpoint.	Keyword: 日 (Sun) Volume: 1 Ref Word: 天部上 (Heaven) Content: “《易》曰：日月丽乎天。又曰：离为日。又曰：日中则昃，月盈则食。天地盈虚，与时消息。而况于人乎，况于鬼神乎…”
《初学记》 Chuxueji	Compiled by Xu Jian (Tang). An introductory encyclopedia strictly organized by chapters (heaven, earth, etc.), drawing from historical literature and preserving ancient fragments.	Keyword: 日 (Sun) Volume: 1 Section: 天部(Heaven) Content: 叙事(Narrative): "说文云日者实也…"; 事对(Events): 丽天&出地（易曰日月丽乎天百谷草木丽乎地文子曰日出于地…; 诗文(Poetry): "梁简文帝咏朝日诗（团团出天外煜煜上层峰光随浪高下影逐树轻浓）…"
《骈字类编》 Pianzi Leibian	Compiled by Shen Zongjing et al. under imperial order (Qing). It collects disyllabic compound words ( Pianzi ) with sources for phrasing reference.	Keyword: 天地 (Heaven and Earth) Volume: 1 Section: 天 (Heaven) Content: “易干夫大人者与天地合其德，又坤天地变化草木蕃天地闭贤人隐…”

Appendix

Table 5: Overview and data samples of the 11 utilized classical sourcebooks.

Human-AI Co-Creation

Human-AI co-creation studies how people and AI jointly develop ideas, artifacts, and decisions while preserving meaningful human agency.

See all 154 papers

Who Wrote This Line? Evaluating the Detection of LLM-Generated Classical Chinese Poetry
Apr 11, 2026Jiang Li, Tian Lan, Shanshan Wang +5AI-Generated Text Detection Computational Literary Studies
College of Computer Science, Inner Mongolia University, China · National & Local Joint Engineering Research Center of Intelligent Information Processing Technology for Mongolian, China · NLP2CT Lab, Department of Computer and Information Science, University of Macau, China
Creating a digital poet
Feb 18, 2026Vered Tohar, Tsahi Hayat, Amir LeshemComputational Creativity Text Generation
Department of Jewish Literature, Bar-Ilan University, Ramat-Gan, Israel. · Department of Communication, Reichman University, Hertzeliya, Israel. · Faculty of Engineering, Bar-Ilan University, Ramat-Gan, Israel.
Poller: Are LLMs Suitable for Evaluating the Poetry Understanding Task?
Jun 29, 2026Shanshan Wang, Derek F. Wong, Jingming Yao +1LLM Evaluation LLM-as-a-Judge
NLP2CT Lab, Department of Computer and Information Science, University of Macau · Department of Portuguese, Faculty of Arts and Humanities, University of Macau

Apr 11, 2026J/K move · Enter open · S save

Who Wrote This Line? Evaluating the Detection of LLM-Generated Classical Chinese Poetry

Jiang Li, Tian Lan, Shanshan Wang, Dongxing Zhang+4

College of Computer Science, Inner Mongolia University, China · National & Local Joint Engineering Research Center of Intelligent Information Processing Technology for Mongolian, China · NLP2CT Lab, Department of Computer and Information Science, University of Macau, China

AI-Generated Text Detection Computational Literary Studies

Open paper AskPDF arXiv

The rapid development of large language models (LLMs) has extended text generation tasks into the literary domain. However, AI-generated literary creations has raised increasingly prominent issues of creative authenticity and ethics in literary world, making the detection of LLM-generated literary texts essential and urgent. While previous works have made significant progress in detecting AI-generated text, it has yet to address classical Chinese poetry. Due to the unique linguistic features of classical Chinese poetry, such as strict metrical regularity, a shared system of poetic imagery, and flexible syntax, distinguishing whether a poem is authored by AI presents a substantial challenge. To address these issues, we introduce ChangAn, a benchmark for detecting LLM-generated classical Chinese poetry that containing total 30,664 poems, 10,276 are human-written poems and 20,388 poems are generated by four popular LLMs. Based on ChangAn, we conducted a systematic evaluation of 12 AI detectors, investigating their performance variations across different text granularities and generation strategies. Our findings highlight the limitations of current Chinese text detectors, which fail to serve as reliable tools for detecting LLM-generated classical Chinese poetry. These results validate the effectiveness and necessity of our proposed ChangAn benchmark. Our dataset and code are available at https://github.com/VelikayaScarlet/ChangAn.

Abstract

Human-AI Co-Creation

Explore similar work

Who Wrote This Line? Evaluating the Detection of LLM-Generated Classical Chinese Poetry

Creating a digital poet

Poller: Are LLMs Suitable for Evaluating the Poetry Understanding Task?