Pith. sign in

REVIEW 3 major objections 6 minor 51 references

Classic4Children: Adapting Chinese Literary Classics for Children with Large Language Model

T0 review · 3 major / 6 minor · reviewed 2026-08-09 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read InstructChild adapts Chinese literary classics into child-friendly text by injecting character personality, narrative structure, and a readability reward into an LLM, and reports higher scores than GPT-4o on the new Classic4Children…

desk verdict A genuinely useful new dataset and task, but the headline readability gains are likely an artifact of optimizing the evaluation metric; the paper needs an independent evaluation before its claims hold. read the letter →

arxiv 2502.01090 v1 pith:YLOB637H submitted 2025-02-03 cs.CL cs.AI

Jiali Chen , Xusen Hei , Yuqi Xue , Zihan Wu , Jiayuan Xie , Yi Cai This is my paper · ORCID
classification cs.CLcs.AI
keywords child-friendlyliteraryadaptationChineseclassicsreadabilitymetricinstructiontuningdirectpreferenceoptimizationlookaheaddecodingClassic4Childrenlargelanguagemodels
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Chinese literary classics are culturally central but are written in classical Chinese with dense plots, so children struggle with them. This paper proposes a child-friendly literary adaptation (CLA) task: automatically rewriting such classics into text that is vivid, concise, and readable for children. The authors build Classic4Children, a parallel dataset of original and expert-adapted passages from the Four Great Classical Novels, and propose InstructChild, which fine-tunes an LLM with character personalities and narrative-structure triples, then uses a readability metric (Red-CN) as a reward in DPO refinement and as a guide during decoding. On the dataset, InstructChild reaches a Red-CN score of 73.12, compared with 68.19 for GPT-4o, and human raters judge it higher on character clarity and narrative conciseness. The paper's core claim is that explicit modeling of children's reading preferences, not larger models or generic simplification, is what moves the needle.

What carries the argument

The load-bearing object is the Red-CN readability metric, a weighted composite that maps an adapted sentence to a reward: $\hat{r} = 0.3\hat{r}_{ac} + 0.4\hat{r}_f + 0.3\hat{r}_t$, where $\hat{r}_{ac}$ is a Gaussian-normalized score for the proportion of adverbs and conjunctions centered at 5, $\hat{r}_f$ is a Gaussian-normalized score for character frequency centered at 85, and $\hat{r}_t$ rewards output brevity relative to the input. This single number plays three roles: it defines what 'child-friendly readability' means, it is the reward signal for DPO refinement, and it is the lookahead score used to rerank candidate tokens during decoding. The other carrier is the integrative instruction itself—personality scores and narrative-structure triples prepended to the input—which steers the model's content decisions, while Red-CN steers its style decisions.

What would settle it

Have children and primary-school teachers rank a set of human-written and model-adapted passages by how easy they are to understand, then compute Red-CN on the same passages; the central claim is falsified if Red-CN does not rank-order the texts the way the human readers do, because the metric is both the training reward and the main evaluation.

Watch

Extended reading notes

Core claim

InstructChild is a three-stage pipeline. First, fine-grained instruction tuning prepends to each original passage a prompt containing Big Five personality scores and brief descriptions of the characters in the passage, together with entity-relation triples extracted from the narrative, so the frozen Qwen2-7B-Instruct backbone learns to generate text that foregrounds personality and follows a concise storyline. Second, a dedicated Chinese readability metric, Red-CN, is computed from the proportion of adverbs and conjunctions, the frequency of characters, and the relative output length, weighted as 0.3, 0.4, and 0.3; this score is used as the reward in direct preference optimization on roughly 1,000 sampled training instances. Third, during inference a lookahead decoding strategy generates several candidate continuations, scores them with Red-CN, and biases token selection toward higher-scoring futures. The authors claim that this combination significantly outperforms closed-source and open-source LLMs, style-transfer models, and fine-tuned baselines on Classic4Children, with the largest gains on Red-CN and on human-judged character clarity and narrative efficiency.

Load-bearing premise

The load-bearing premise is that the Red-CN readability score—a combination of adverb/conjunction proportion, character frequency, and output length—truly measures how easy a Chinese text is for a child, because that same score is used as the training reward and as the headline evaluation metric, and its target values are taken from the expert-adapted texts used for training.

Editorial extensions

If this is right

  • If InstructChild works as reported, child-friendly adaptation of Chinese classics can be automated at paragraph scale, giving teachers and publishers a starting draft that still requires human review.
  • The Classic4Children dataset, with 2,686 training and 300 test paragraph pairs across the Four Great Classical Novels, provides a reusable benchmark for future CLA systems.
  • The Red-CN metric, used both as a training reward and as a decoding guide, correlates with human judgments on fluency, content preservation, character clarity, and narrative efficiency, so it can serve as a cheap proxy during development.
  • The three components contribute unevenly: removing personality or narrative structure lowers Red-CN, removing DPO refinement drops Red-CN more sharply, and removing lookahead decoding causes a smaller loss.
  • Because larger API models like GPT-4o and GLM-4 produced longer, more detailed output, the results imply that prompting alone is insufficient for this task; explicit preference modeling is the differentiator.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • If Red-CN's target values (5 for adverb/conjunction proportion, 85 for character frequency) were re-estimated on a broader corpus of child literature, the metric might transfer beyond the Four Great Classical Novels; the paper currently derives them from the same expert-adapted editions it trains and evaluates on.
  • The same recipe—personality prompts plus a readability reward plus lookahead decoding—could be tested on other culturally dense heritage texts, such as classical poetry or historical narratives, where the bottleneck is also classical-language complexity rather than vocabulary alone.
  • A direct test of the metric's generality would be to rebuild the pipeline with Red-CN targets estimated from a held-out set of children's books that are not part of Classic4Children; if Red-CN scores still track human readability judgments, the claim is materially stronger.
  • Because the paper's case study shows that simplification can introduce misunderstandings (the sword reaching Lü Bu), a production version would need a content-fidelity check to prevent plot errors; the paper acknowledges this as future work.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 6 minor

Summary. The paper introduces the child-friendly literary adaptation (CLA) task for Chinese literary classics and constructs the Classic4Children dataset from the Four Great Classical Novels, pairing original passages with human-adapted children's versions. The proposed method, InstructChild, augments a Qwen2-7B-Instruct backbone with three components: fine-grained instruction tuning that incorporates character personality scores and entity-relation narrative structure, a DPO refinement stage that uses a newly designed Chinese readability metric (Red-CN) as the reward, and a lookahead decoding strategy that reranks candidate continuations by Red-CN at inference. Experiments on Classic4Children compare InstructChild against API-based LLMs, text style transfer models, and fine-tuned open-source LLMs, reporting gains in BLEU, BERTScore, Red-CN, and a small human evaluation.

Significance. If the central claim were established, the paper would make a useful contribution: it defines a concrete task, releases a new benchmark dataset built from real published child adaptations, and demonstrates a training-plus-inference recipe that integrates readability signals into an LLM. The ablation study and hyperparameter sensitivity tables are constructive, and the authors are transparent about computational costs and limitations of the lookahead strategy. However, the headline readability improvement is currently not credible as evidence for child-appropriate text, because Red-CN is simultaneously the training reward, the decoding guidance, and the main automatic evaluation metric, with its target values and weights derived from the same expert-adapted corpus. The human evaluation is too small and too indirect to break the circularity. The central claim therefore needs substantial re-validation before the paper can be accepted.

major comments (3)
  1. [§3.2.1, Eq. (2)–(5); §3.2.2; §3.3; Table 1] The main evaluation metric Red-CN is constructed from the same corpus that defines the training objective, making the reported headline gains partly circular. Specifically, Eqs. (2)–(5) define Red-CN with targets rac=5, rf=85, and weights (0.3, 0.4, 0.3) chosen from the authors' analysis of the expert-adapted texts; §3.2.2 then uses Red-CN as the DPO reward, and Eq. (7) uses it as the lookahead decoding guidance. Reporting Red-CN as the headline metric in Table 1 therefore largely measures how well the model optimizes its own reward, not whether children can read the output. The concern is concrete: StoryTrans and ParaGuide achieve Red-CN scores of 69.74 and 71.87 with much lower BLEU scores (7.59 and 13.88), which is consistent with Red-CN being gameable by short, content-poor output through the length term in Eq. (4). The paper needs an independent readability validation, such as a held-out readability assessment using an external readability formula or graded readers, and it should avoid claiming that a Red-CN gain alone constitutes a significant improvement in children's readability.
  2. [§4.4.2, §4.5.3, Table 2] The human evaluation is too limited to establish that InstructChild produces text that is actually more readable for children. Only 200 test samples and five adult volunteers with Chinese educational backgrounds participated; no children were involved, and no significance tests, inter-annotator agreement measures, or per-criterion statistical comparisons are reported. The reported standard deviations are descriptive only and do not support the claim that the model 'significantly' outperforms baselines on human evaluation. Moreover, InstructChild obtains a lower Content Preservation score than GPT-4o (1.69 vs. 1.85), and §4.6 acknowledges that its simplification can cause misunderstandings. The human evaluation therefore does not yet counterbalance the circularity of the automatic Red-CN metric. I would like to see a larger human study with child participants (or reading-comprehension tests), explicit significance testing, and reporting of content preservation failures.
  3. [§4.5.1, Table 1] The paper uses the word 'significantly' throughout the abstract, introduction, and conclusions, but no statistical significance tests are reported for the automatic metrics. Since the baseline differences in Table 1 are often small (e.g., 73.12 vs. 72.97 for InstructChild vs. InstructChild w/o Look, or 70.21 vs. 69.94 for BS-F1), the claim of 'significant performance gains' needs either paired significance tests over the 300 test samples or explicit reporting of variance across seeds. Without such tests, the improvements should be described as numerical gains.
minor comments (6)
  1. [§2.1] There is a typo in 'promt-based editing'; it should be 'prompt-based editing'.
  2. [§1, Figure 1] The caption of Figure 1 includes a fragment 'Evenlonger than originaltext' that appears to be a leftover annotation; it should be integrated into the caption or removed.
  3. [§4.5.3, Table 2] In Table 2, the entries for GPT-4o's NE column appear as '0.95/0.12' with no visible separator issue, but the printed value '1.95/0.25' for Flu is slightly cramped; more importantly, the table would be clearer if the metric value and standard deviation were labeled in the caption rather than defined only in the table footnote.
  4. [§3.2.1, Eq. (2)] The standard deviation δ is introduced in Eq. (2) but its value is only given later in §4.2; moving the parameter values to the first mention would help readers follow the metric definition.
  5. [§1] The sentence 'the adaptation also take into account' has a subject-verb agreement error; it should read 'the adaptation also takes into account'.
  6. [Appendix B.2, Table 6] The correlation coefficients in Table 6 are reported without the number of samples or confidence intervals; since the table is used to argue that Red-CN correlates with human judgments, adding these details and clarifying that the 100 samples are separate from the 200-sample human evaluation would strengthen the presentation.

Circularity Check

1 steps flagged · score 6.0 of 10

Red-CN serves as both the DPO/lookahead optimization objective and the headline evaluation metric, so the claimed readability gain is partly forced by the metric itself rather than independently measured.

  1. fitted input called prediction [Sec. 3.2.1 (Eqs. 2-5), Sec. 3.2.2, Sec. 3.3 (Eqs. 6-7), Sec. 4.4.1, Table 1]
    "We analyze expert-adapted child-friendly literary classics and find that these two indicators typically cluster around values of 5 and 85 per sentence, respectively. ... we assign the specific weights to obtain the overall readability metric (i.e., Red-CN), as the reward score ˆr: ˆr = 0.3ˆrac + 0.4ˆrf + 0.3ˆrt. ... we also use the readability scores in Eq. 5 as an additional metric (i.e., Red-CN) to assess whether the generated text aligns with children’s reading levels (higher is better)."

    Red-CN is defined (Eqs. 2-5) with target values 5 and 85 taken from the same expert-adapted texts used to build Classic4Children, and it is used as the DPO ranking reward (Sec. 3.2.2) and as the lookahead-decoding objective G(·)=ˆr (Eqs. 6-7). Section 4.4.1 then reports this same Eq. 5 score as the headline Red-CN metric in Table 1. InstructChild's higher Red-CN (73.12 vs GPT-4o's 68.19) is therefore partly forced by the training/inference objective. BLEU/BERTScore measure similarity to expert references, not readability; the human study has no child participants and reports lower Content Preservation for InstructChild, which the authors say 'can sometimes lead to misunderstandings' (Sec. 4.6). The headline readability gain is a fitted objective reported as a prediction.

full rationale

Most of the pipeline (personality annotation, narrative triplets, instruction tuning, DPO, lookahead) is not circular: BFPT-based personality and entity-relation extraction are independent auxiliary inputs, and BLEU/BERTScore are reference-based similarity measures that are not used as training objectives. However, the paper's central readability claim is circular in a narrow but load-bearing way: Red-CN (Eq. 5) is constructed from expert-adapted texts (targets rac=5, rf=85), used as the reward for DPO ranking and as the lookahead-decoding selection objective (Eqs. 6-7), and then reported as the main evaluation metric in Table 1. This makes the headline Red-CN advantage partially an artifact of optimizing the evaluation metric itself. The correlation analysis in Table 6 is an attempt at external validation, but it is computed on generated samples and does not break the reward/evaluation identity. The human evaluation (higher CC, NE) gives some independent support, but it is not a child readability test and InstructChild scores lower on Content Preservation, which the authors acknowledge can cause misunderstanding. Overall the non-Red-CN results keep the paper from being wholly circular, but the chief 'significant readability gain' is a fitted objective being reported as a prediction.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The central claim relies on Red-CN, a metric whose hyperparameters are fitted to the same expert-adapted corpus used for training and evaluation, plus assumptions about GPT-4o annotations and child reading preferences that are not externally validated.

free parameters (5)
  • Red-CN target for adverb/conjunction proportion rac = 5
    Set to the value expert-adapted child-friendly classics cluster around, as stated in Section 3.2.1.
  • Red-CN target for character frequency rf = 85
    Set to the value expert-adapted child-friendly classics cluster around, as stated in Section 3.2.1.
  • Gaussian standard deviations delta = 2.5 and 42.5
    Chosen as half the expectation values, stated in Section 4.2 as a manual setting.
  • Red-CN weights (w_rac, w_rf, w_t) = 0.3, 0.4, 0.3
    Assigned in Eq. 5 following Da (2004), with no optimization or external validation described.
  • Lookahead hyperparameters L, n, lambda = 5, 20, 1
    Selected based on test-set sensitivity experiments in Appendix B.3, so tuned to the test set.
assumptions (4)
  • domain assumption GPT-4o's Big Five personality scores for characters are accurate and useful for generating child-friendly text.
    The method prepends these scores to fine-tuning instructions, assuming they help produce vivid character portrayals (Section 3.1.1).
  • domain assumption Entity-relation triplets extracted by GPT-4o capture the concise narrative structure needed for children.
    The narrative structure is represented solely by these triplets, assuming they help simplify plots (Section 3.1.2).
  • domain assumption The three identified children's reading preferences (vivid characters, concise narrative, appropriate readability) are correct and universal.
    These preferences are inferred from a single human-written adaptation example and are not validated with child readers or child development literature (Section 1, Figure 1).
  • ad hoc to paper Red-CN is a valid proxy for children's readability.
    The metric is defined with targets fitted to the same corpus and used as both training reward and evaluation metric; validity rests only on a 100-sample correlation with volunteer judgments (Section 3.2.1, Appendix B.2).

how reviews work

0 comments
Cite this review

Pith. "Pith review of Classic4Children: Adapting Chinese Literary Classics for Children with Large Language Model." pith.science (2026). https://pith.science/paper/YLOB637H

@misc{pith2026250201090,
  author       = {Pith},
  title        = {Pith review of: Classic4Children: Adapting Chinese Literary Classics for Children with Large Language Model},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/YLOB637H}},
  note         = {Machine review of arXiv:2502.01090}
}
read the original abstract

Chinese literary classics hold significant cultural and educational value, offering deep insights into morality, history, and human nature. These works often include classical Chinese and complex narratives, making them difficult for children to read. To bridge this gap, we introduce a child-friendly literary adaptation (CLA) task to adapt the Chinese literary classic into engaging and accessible text for children. However, recent large language models (LLMs) overlook children's reading preferences (\ie, vivid character portrayals, concise narrative structures, and appropriate readability), which poses challenges in CLA. In this paper, we propose a method called InstructChild, which augments the LLM with these preferences for adaptation. Specifically, we first obtain the characters' personalities and narrative structure as additional information for fine-grained instruction tuning. Then, we devise a readability metric as the reward to align the LLM with the children's reading level. Finally, a lookahead decoding strategy is applied to improve the readability of the generated text during inference. To support the evaluation of CLA task, we construct the Classic4Children dataset, which comprises both the original and child-friendly versions of the Four Great Classical Novels of Chinese literature. Experimental results show that our InstructChild significantly improves automatic and human evaluation performance.

Figures

Figures reproduced from arXiv: 2502.01090 by the authors.

Figure 1
Figure 1. The sample from our Classic4Children dataset. [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Overview of InstructChild. (a) fine-grained instruction tuning, which incorporates the characters’ [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. shows the adapted text generated by GPT-4o and our InstructChild. Specifically, InstructChild produces a more concise narrative that emphasizes key characters’ traits. It describes Cao Cao’s deter￾mination and urgency while simplifying his inner monologue. In contrast, GPT-4o solely rewrites the original text, resulting in overly lengthy con￾tent with complex words. Moreover, the output of GPT-4o is even longer than… view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Some cases of personality assessment. L BLEU-2 BS-F1 Red-CN 2 12.86 69.98 73.01 5 13.09 70.21 73.12 8 13.27 70.67 73.15 [PITH_FULL_IMAGE:figures/full_fig_p012_4.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

51 extracted references · 39 canonical work pages

  1. [1]

    Jiali Chen, Zhenjun Guo, Jiayuan Xie, Yi Cai, and Qing Li. 2023. Deconfounded visual question generation with causal inference. In Proc. of ACM MM, pages 5132--5142. ACM

    work page 2023

  2. [2]

    Jiali Chen, Xusen Hei, Yuqi Xue, Yuancheng Wei, Jiayuan Xie, Yi Cai, and Qing Li. 2024. Learning to correction: Explainable feedback generation for visual commonsense reasoning distractor. In Proc. of ACM MM, pages 8209--8218. ACM

    work page 2024

  3. [3]

    Jianlin Chen. 2024. Lmstyle benchmark: Evaluating text style transfer for chatbots. CoRR, abs/2403.08943

    arXiv 2024

  4. [4]

    Madalina Chitez, Mihai Dascalu, Aura Cristina Udrea, Cosmin Striletchi, Karla Cs \" u r \" o s, Roxana Rogobete, and Alexandru Oravitan. 2024. Towards building the LEMI readability platform for children's literature in the romanian language. In Proc. of COLING, pages 16450--16456. ELRA and ICCL

    work page 2024

  5. [5]

    Yiming Cui, Ziqing Yang, and Xin Yao. 2023. Efficient and effective text encoding for chinese llama and alpaca. CoRR, abs/2304.08177

    arXiv 2023

  6. [6]

    Jun Da. 2004. A corpus-based study of character and bigram frequencies in chinese e-texts and its implications for chinese language instruction. In Proceedings of the fourth international conference on new technologies in teaching and learning Chinese, pages 501--511. Citeseer

    work page 2004

  7. [7]

    Daryna Dementieva, Daniil Moskovskiy, David Dale, and Alexander Panchenko. 2023. Exploring methods for cross-lingual text style transfer: The case of text detoxification. In Proc. of IJCNLP, pages 1083--1101. Association for Computational Linguistics

    work page 2023

  8. [8]

    Qingxiu Dong, Lei Li, Damai Dai, Ce Zheng, Zhiyong Wu, Baobao Chang, Xu Sun, Jingjing Xu, and Zhifang Sui. 2022. A survey on in-context learning. arXiv preprint arXiv:2301.00234

    arXiv 2022

Show all 51 references
  1. [9]

    Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. 2024. The llama 3 herd of models. arXiv preprint arXiv:2407.21783

    2024 arXiv

  2. [10]

    Team GLM, Aohan Zeng, Bin Xu, Bowen Wang, Chenhui Zhang, Da Yin, Diego Rojas, Guanyu Feng, Hanlin Zhao, Hanyu Lai, et al. 2024. Chatglm: A family of large language models from glm-130b to glm-4 all tools. arXiv preprint arXiv:2406.12793

    2024 arXiv

  3. [11]

    Zachary Horvitz, Ajay Patel, Chris Callison - Burch, Zhou Yu, and Kathleen R. McKeown. 2024. Paraguide: Guided diffusion paraphrasers for plug-and-play textual style transfer. In Proc. of AAAI, pages 18216--18224. AAAI Press

    2024

  4. [12]

    Chih-tsing Hsia. 2016. The classic Chinese novel: A critical introduction. The Chinese University of Hong Kong Press

    2016

  5. [13]

    Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen - Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen

    Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen - Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. 2022. Lora: Low-rank adaptation of large language models. In Proc. of ICLR. OpenReview.net

    2022

  6. [14]

    Zhiyuan Hu, Yuliang Liu, Jinman Zhao, Suyuchen Wang, Yan Wang, Wei Shen, Qing Gu, Anh Tuan Luu, See - Kiong Ng, Zhiwei Jiang, and Bryan Hooi. 2024. Longrecipe: Recipe for efficient long context generalization in large language models. CoRR, abs/2409.00509

    2024 arXiv

  7. [15]

    Haifeng Hui. 2024. Children's literature adaptation studies: Interdisciplinary theoretical resources with a special focus for education. History of Education and Children's Literature: HECL: XIX, 1, 2024, pages 629--647

    2024

  8. [16]

    Tannon Kew, Alison Chi, Laura V \' a squez - Rodr \' guez, Sweta Agrawal, Dennis Aumiller, Fernando Alva - Manchego, and Matthew Shardlow. 2023. BLESS: benchmarking large language models on sentence simplification. In Proc. of EMNLP, pages 13291--13309. Association for Computa...

    2023

  9. [17]

    Dhruv Kumar, Lili Mou, Lukasz Golab, and Olga Vechtomova. 2020. Iterative edit-based unsupervised sentence simplification. In Proc. of ACL, pages 7918--7928. Association for Computational Linguistics

    2020

  10. [18]

    James Legge. 2022. The Chinese Classics: Vol. I. BoD--Books on Demand

    2022

  11. [19]

    Pusheng Liu, Lianwei Wu, Linyong Wang, Sensen Guo, and Yang Liu. 2024 a . Step-by-step: Controlling arbitrary style in text with large language models. In Proc. of COLING, pages 15285--15295. ELRA and ICCL

    2024

  12. [20]

    Qingyi Liu, Jinghui Qin, Wenxuan Ye, Hao Mou, Yuxuan He, and Keze Wang. 2024 b . Adaptive prompt routing for arbitrary text style transfer with pre-trained language models. In Proc. of AAAI, pages 18689--18697. AAAI Press

    2024

  13. [21]

    Zhao Luo. 2019. Research on the teaching of chinese language and literature in modern educational thoughts. Review of Educational Theory, 2(2):29--33

    2019

  14. [22]

    Huiyu Mai, Wenhao Jiang, and Zhi - Hong Deng. 2023. Prefix-tuning based unsupervised text style transfer. In Proc. of EMNLP Findings, pages 14847--14856. Association for Computational Linguistics

    2023

  15. [23]

    Jonathan Mallinson, Aliaksei Severyn, Eric Malmi, and Guillermo Garrido. 2020. FELIX: flexible text editing through tagging and insertion. In Proc. of EMNLP Findings, volume EMNLP 2020 of Findings of ACL , pages 1244--1255. Association for Computational Linguistics

    2020

  16. [24]

    Ojha, and Ondrej Dusek

    Sourabrata Mukherjee, Atul Kr. Ojha, and Ondrej Dusek. 2024. Are large language models actually good at text style transfer? In Proc. of INLG, pages 523--539. Association for Computational Linguistics

    2024

  17. [25]

    Anja M \"u ller. 2013. Adapting Canonical Texts in Children's Literature. A&C Black

    2013

  18. [26]

    OpenAI. 2023. GPT-4 technical report. CoRR, abs/2303.08774

    2023 arXiv

  19. [27]

    Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll L. Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, John Schulman, Jacob Hilton, Fraser Kelton, Luke Miller, Maddie Simens, Amanda Askell, Peter Welinder, Paul F. Christiano, Jan Leik...

    2022

  20. [28]

    Kishore Papineni, Salim Roukos, Todd Ward, and Wei - Jing Zhu. 2002. Bleu: a method for automatic evaluation of machine translation. In Proc. of ACL, pages 311--318

    2002

  21. [29]

    Jipeng Qiang, Yun Li, Yi Zhu, Yunhao Yuan, and Xindong Wu. 2020. Lsbert: A simple framework for lexical simplification. CoRR, abs/2006.14939

    2020 arXiv

  22. [30]

    Manning, Stefano Ermon, and Chelsea Finn

    Rafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D. Manning, Stefano Ermon, and Chelsea Finn. 2023. Direct preference optimization: Your language model is secretly a reward model. In Proc. of NeurIPS

    2023

  23. [31]

    Emily Reif, Daphne Ippolito, Ann Yuan, Andy Coenen, Chris Callison - Burch, and Jason Wei. 2022. A recipe for arbitrary text style transfer with large language models. In Proc. of ACL, pages 837--848. Association for Computational Linguistics

    2022

  24. [32]

    Sonia Roccas, Lilach Sagiv, Shalom H Schwartz, and Ariel Knafo. 2002. The big five personality factors and personal values. Personality and social psychology bulletin, 28(6):789--801

    2002

  25. [33]

    Mirac Suzgun, Luke Melas - Kyriazi, and Dan Jurafsky. 2022. Prompt-and-rerank: A method for zero-shot and few-shot arbitrary textual style transfer with small language models. In Proc. of EMNLP, pages 2195--2222. Association for Computational Linguistics

    2022

  26. [34]

    Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, Dan Bikel, Lukas Blecher, Cristian Canton - Ferrer, Moya Chen, Guillem Cucurull, David Esiobu, Jude Fernandes, Jeremy Fu,...

    2023 arXiv

  27. [35]

    Maria Valentini, Jennifer Weber, Jesus Salcido, T \' e a Wright, Eliana Colunga, and Katharina von der Wense. 2023. On the automatic generation and simplification of children's stories. In Proc. of EMNLP, pages 3588--3598. Association for Computational Linguistics

    2023

  28. [36]

    Peifeng Wang, Zhengyang Wang, Zheng Li, Yifan Gao, Bing Yin, and Xiang Ren. 2023. SCOTT: self-consistent chain-of-thought distillation. In Proc. of ACL, pages 5546--5558. Association for Computational Linguistics

    2023

  29. [37]

    Wong, Jingming Yao, and Lidia S

    Shanshan Wang, Derek F. Wong, Jingming Yao, and Lidia S. Chao. 2024 a . What is the best way for chatgpt to translate poetry? CoRR, abs/2406.03450

    2024 arXiv

  30. [38]

    Xintao Wang, Yunze Xiao, Jen - tse Huang, Siyu Yuan, Rui Xu, Haoran Guo, Quan Tu, Yaying Fei, Ziang Leng, Wei Wang, Jiangjie Chen, Cheng Li, and Yanghua Xiao. 2024 b . Incharacter: Evaluating personality fidelity in role-playing agents through psychological interviews. In Proc...

    2024

  31. [39]

    Jiayuan Xie, Yi Cai, Jiali Chen, Ruohang Xu, Jiexin Wang, and Qing Li. 2024. Knowledge-augmented visual question answering with natural language explanation. IEEE Trans. Image Process. , 33:2652--2664

    2024

  32. [40]

    Wei Xu, Courtney Napoles, Ellie Pavlick, Quanze Chen, and Chris Callison - Burch. 2016. Optimizing statistical machine translation for text simplification. Trans. Assoc. Comput. Linguistics, 4:401--415

    2016

  33. [41]

    Wenda Xu, Danqing Wang, Liangming Pan, Zhenqiao Song, Markus Freitag, William Wang, and Lei Li. 2023. INSTRUCTSCORE: towards explainable text generation evaluation with automatic feedback. In Proc. of EMNLP, pages 5967--5994. Association for Computational Linguistics

    2023

  34. [42]

    An Yang, Baosong Yang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Zhou, Chengpeng Li, Chengyuan Li, Dayiheng Liu, Fei Huang, et al. 2024. Qwen2 technical report. arXiv preprint arXiv:2407.10671

    2024 arXiv

  35. [43]

    Li Yuan, Yi Cai, Haopeng Ren, and Jiexin Wang. 2024. A logical pattern memory pre-trained model for entailment tree generation. In Proc. of COLING, pages 759--772

    2024

  36. [44]

    Li Yuan, Jin Wang, Liang-Chih Yu, and Xuejie Zhang. 2022. Hierarchical template transformer for fine-grained sentiment controllable generation. Information Processing & Management, 59(5):103048

    2022

  37. [45]

    Weinberger, and Yoav Artzi

    Tianyi Zhang, Varsha Kishore, Felix Wu, Kilian Q. Weinberger, and Yoav Artzi. 2020. Bertscore: Evaluating text generation with BERT . In Proc. of ICLR. OpenReview.net

    2020

  38. [46]

    Jie Zhao, Ziyu Guan, Cai Xu, Wei Zhao, and Yue Jiang. 2024 a . SC2: towards enhancing content preservation and style consistency in long text style transfer. CoRR, abs/2406.04578

    2024 arXiv

  39. [47]

    Jinman Zhao, Zifan Qian, Linbo Cao, Yining Wang, and Yitian Ding. 2024 b . Bias and toxicity in role-play reasoning. CoRR, abs/2409.13979

    2024 arXiv

  40. [48]

    Jinman Zhao and Xueyan Zhang. 2024. Large language model is not a (multilingual) compositional relation reasoner. In First Conference on Language Modeling

    2024

  41. [49]

    Xuekai Zhu, Jian Guan, Minlie Huang, and Juan Liu. 2023. Storytrans: Non-parallel story author-style transfer with discourse representations and content enhancing. In Proc. of ACL, pages 14803--14819. Association for Computational Linguistics

    2023

  42. [50]

    online" 'onlinestring :=

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint eprinttype howpublished institution journal key month note number organization pages publisher school series title type volume year doi pubmed url lastchecked label extra.label sort.label short.list...

  43. [51]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 gl...

Pith tools

Reviewed August 9, 2026 · model on record in the stance chip above.