Pith. sign in

REVIEW 4 major objections 6 minor 3 cited by

Redefining Simplicity: Benchmarking Large Language Models from Lexical to Document Simplification

T0 review · 4 major / 6 minor · reviewed 2026-08-08 · deepseek-v4-flash

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

Pith's one-line read Large language models outperform dedicated non-LLM simplification systems on all four text-simplification tasks, and GPT-4o output is often judged better than the human-written references.

desk verdict Useful cross-task benchmark with an overstated headline: the human-eval evidence for 'better than human references' is too weak, and 'all four tasks' depends on which metric you read. read the letter →

arxiv 2502.08281 v1 pith:3WPOBEU5 submitted 2025-02-12 cs.CL

Jipeng Qiang , Minjiang Huang , Yi Zhu , Yunhao Yuan , Chaowei Zhang , Kui Yu This is my paper · ORCID
classification cs.CL
keywords textsimplificationlargelanguagemodelslexicalsyntacticsentencedocumentbenchmarkinghumanevaluation
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

This paper tries to establish that general-purpose large language models have overtaken dedicated text-simplification systems at every level of the task, from replacing single hard words to rewriting whole documents. The authors benchmark GPT-4o, Llama3.1-70B, and the lightweight Gemma2-2B against the best published non-LLM baselines on four standard tasks, and report that the LLMs win on automatic metrics across the board. They also report human evaluations in which GPT-4o's output was judged better than the dataset's own human-written references in a majority of sampled cases on most aspects. If that is right, supervised simplification models lose their reason to exist, and reference-based evaluation metrics become a ceiling that no longer measures true simplification quality; the field's bottleneck shifts from model capability to evaluation design and harder task formulations.

What carries the argument

The load-bearing mechanism is the four-task benchmark design: for each task the paper fixes a standard dataset and metric suite, then varies the model class while holding the prompt and evaluation procedure constant. Lexical simplification uses TSAR-2022 with ACC@1, MAP@k, and Potential@k; sentence simplification uses ASSET and MED-EASI with SARI, BERTScore, FKGL, and LENS; syntactic simplification uses WEBSPLIT with BERTScore and L2SCA indices; document simplification uses Newsela with SARI, D-SARI, BARTScore, and FKGL. A side-by-side human evaluation in which annotators choose between GPT-4o output and the human reference supplies the evidence that LLM output can exceed the reference quality.

What would settle it

Recompute every quoted baseline score with the same metric implementations, the same test splits, and the same decoding setup used for the LLM outputs; if the re-run non-LLM scores match or exceed the quoted LLM scores on any of the four tasks, the paper's across-the-board superiority claim would be shown to be an artifact of comparing incomparable numbers. A second decisive check is a preregistered human study where professional editors, rather than graduate students, blind-compare GPT-4o output against human references on the same 100-instance samples.

Watch

Extended reading notes

Core claim

On the paper's own terms, the discovery is a complete inversion of the previous ranking in text simplification: prompted general-purpose LLMs, especially GPT-4o, produce outputs that score higher than specialized supervised models on lexical, syntactic, sentence, and document simplification. The human evaluation goes further, finding that GPT-4o output is preferred over the human-annotated reference in most sampled cases for overall quality in all four tasks, with the strongest advantage in syntactic simplification. This makes the human references a floor rather than a ceiling, and it implies that the remaining bottlenecks in text simplification are not basic model capability but evaluation design, cost, and harder task formulations.

Load-bearing premise

The ranking assumes that the older numbers quoted for non-LLM systems were produced on the same data with the same metric code as the new LLM scores, even though those numbers come from earlier papers with different implementations and possibly different splits.

Editorial extensions

If this is right

  • If the central claim is correct, fine-tuned, task-specific simplification models no longer offer a competitive advantage: a prompted general LLM beats them on all four tasks, so new systems should be built around LLMs rather than supervised sequence-to-sequence training.
  • Human-annotated references can no longer serve as an implicit quality ceiling; metrics that reward similarity to those references will systematically undervalue LLM output, so new reference-free or LLM-based metrics are needed.
  • The lightweight 2B model's strong sentence-level and syntactic results with weak lexical and document results suggest small models are viable for low-cost syntactic simplification but not for tasks requiring broader knowledge.
  • Benchmark datasets designed in the pre-LLM era need to be refreshed or replaced, because the upper range of quality they were built to measure has been passed.
  • Future simplification research should concentrate on multi-level and personalized simplification, because standard single-level simplification is largely solved for current LLMs.

Reading between the lines

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

  • Editorial inference: if the cross-paper score comparison is not apples-to-apples, the headline ranking could shrink or reverse; the strongest test is a single-codebase rerun of every baseline on identical data and metric versions.
  • Editorial inference: 'better than human references' may say more about the crowdworker reference standard than about absolute quality; a fairer test would compare LLM output against professional editors or against readers with the target literacy levels.
  • Editorial inference: a cheap, testable extension of the paper is to run the same prompts with newer LLMs and with non-LLM models retrained on LLM-generated simplification pairs, to see whether distillation closes the gap that direct prompting opens.
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

4 major / 6 minor

Summary. This paper benchmarks three large language models (Gemma2-2B, Llama3.1-70B, and GPT-4o) against non-LLM baselines on four text simplification tasks: lexical, syntactic, sentence, and document simplification. The authors report a range of automatic metrics (e.g., SARI, BERTScore, FKGL, LENS, D-SARI, BARTScore, and lexical accuracy) and a human evaluation in which three annotators compare GPT-4o outputs with human references on 100 instances per task. Based on these experiments, they claim that LLMs outperform non-LLM approaches on all four tasks and often generate outputs that exceed the quality of human-annotated references, and they propose four future research directions for text simplification.

Significance. If the headline claims held, this would be a useful unified benchmark, since most prior work has focused on sentence simplification alone. The four-task coverage, the inclusion of lightweight, open-source, and closed-source models, and the release of prompts and experimental results are genuine strengths that support reproducibility. However, the conclusions are broader than the evidence: the human evaluation is underpowered, several baseline numbers are imported from heterogeneous prior work without a common evaluation harness, and the paper's own tables do not uniformly show one model winning every metric. The benchmark itself is a contribution, but the central claims need substantial tempering and additional validation.

major comments (4)
  1. [Section 2.5, Table 6] The human evaluation is too underpowered to support the claim that GPT-4o outputs 'often exceed' human references. It uses only three non-native graduate-student annotators and 100 instances per task, reports no inter-annotator agreement, confidence intervals, or significance tests, and most preference rates are near chance (e.g., LexS Overall 55.3, SenS Simplicity 44.3). Under a simple binomial model with 300 judgments, a 55.3% rate has an approximate 95% confidence interval of about 49.7-60.9%, so the majority preference is not distinguishable from a coin flip. Because the abstract and Section 2.6 rest on this human evidence for the 'exceeds human-annotated references' sub-claim, that sub-claim is not established by the reported data.
  2. [Tables 2-5, Sections 2.1-2.4] Many baseline scores are imported from prior publications rather than computed in this paper's evaluation harness: Table 2 uses Davinci rows from Aumiller and Gertz (2022), Table 3 uses Davinci rows from Kew et al. (2023), Table 4 uses GPT-3.5 and Llama-3 from Yao et al. (2024), and Table 5 uses GPT-3.5 and GPT-4 from Fang et al. (2025). These prior results may involve different metric implementations, preprocessing, model versions, or data splits, as in the 200-document Newsela sample for Table 5. The cross-paper ranking 'LLMs outperform non-LLM approaches in all four tasks' is therefore not apples-to-apples. The authors should either rerun the baselines under identical conditions or restrict the claims to directly comparable evaluations.
  3. [Table 4, Table 5, Section 2.6] The claim in Section 2.6 that 'GPT-4o outperforms all other systems in all tasks using automatic metrics' is contradicted by the paper's own tables. In Table 4, ABCD and GPT-3.5 achieve a mean BERTScore of 0.90 while GPT-4o achieves 0.86, and Llama-3(8B) attains lower MLT and CN/T values. In Table 5, Llama3.1 has the best BARTScore (-2.28) compared with GPT-4o (-2.54). The core claim should identify the specific metrics on which each model wins, and the comparisons should include variance or significance information rather than point estimates alone.
  4. [Section 2.2, Table 3] The MED-EASI part of Table 3 contains a data-quality problem: Davinci-002's FKGL is listed as 43.67, identical to its BERTScore and implausible for a readability grade level, which appears to be a copying error. In addition, the 'Gold References' row reports SARI = 100.00 and BERTScore = 100.00, which are tautological because the references are compared with themselves; this row does not provide a meaningful baseline for the claim that LLM outputs exceed human references in sentence simplification.
minor comments (6)
  1. [Section 2.1, Table 2] The baseline is called 'MANTI' in the text but 'MANTIS' in Table 2; please harmonize the naming.
  2. [Section 1] A sentence in the introduction reads 'existing research only shows that Previous studies have only demonstrated that'; this is duplicated and ungrammatical and should be revised.
  3. [Section 2.2, Metrics] BERTScore is cited to Papineni et al. (2002), which is the BLEU paper; the correct citation is Zhang et al. (2020).
  4. [Section 2.4] The subsection ordering is disordered: '(3) LLMs' appears before '(2) Experimental Setting'; please renumber or reorder the subsections.
  5. [Table 4] The 'Exact Copy' row is not a simplification system and trivially achieves BERTScore 1.00; its role in the comparison should be clarified or the row removed.
  6. [Section 2.5] The paper does not state whether the raw judgments, annotator-level scores, or instance IDs from the human evaluation will be released; without these, the human evaluation cannot be audited or reanalyzed.

Circularity Check

0 steps flagged · score 2.0 of 10

No circular derivation: the benchmark's conclusions rest on empirical measurements, not on equations that reduce to their own inputs.

full rationale

This is an empirical benchmarking paper rather than a derivation, so the self-definitional and fitted-input patterns do not apply. LLMs are not trained or fitted on the test sets; they are queried zero-shot or few-shot with fixed prompts, and their outputs are scored with standard automatic metrics and a separate human preference study. The human study (Section 2.5) is small and statistically underpowered, but weakness of evidence is a validity concern, not circularity. The main comparability risk is that several non-LLM baselines and prompts are imported as reported numbers from the authors' own prior work (LSBERT, LSPG, Feng et al. 2023, Fang et al. 2025). These self-citations are not load-bearing in the circularity sense, because the same rankings are supported by external baselines such as MANTIS, MUSS, KIS, BART-SWIPE, PG, and ABCD, and because no conclusion is defined in terms of those cited results. No uniqueness theorem is invoked, and no ansatz is smuggled in as an external mathematical fact. The paper therefore has no circular step, but the self-cited baseline asymmetry warrants a mild score of 2.

Assumptions & free parameters 0 free parameters · 3 assumptions · 0 invented entities

This is an empirical benchmark, so there are no fitted free parameters and no invented entities. The central claim rests on measurement assumptions: that automatic metrics are valid, that baseline numbers from different papers are comparable, and that the small human evaluation is sufficient to support the strongest wording.

assumptions (3)
  • domain assumption The selected automatic metrics (SARI, BERTScore, FKGL, LENS, D-SARI, BARTScore, L2SCA) are valid proxies for simplification quality and are comparable across systems.
    Used throughout Section 2 to rank systems; if these metrics are noisy or biased, the cross-system ranking may not reflect real simplification quality.
  • domain assumption Baseline scores reported in prior papers are comparable to the new LLM outputs.
    Tables 2 through 5 mix reported baselines with new runs; the paper does not rerun non-LLM baselines under identical conditions.
  • domain assumption The human evaluation protocol (three non-native graduate students, 100 random instances per task, no agreement measure) is representative enough to conclude that LLM outputs exceed human references.
    Section 2.5 and Table 6; the small sample and annotator pool limit the strength of the 'exceed human references' conclusion.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Redefining Simplicity: Benchmarking Large Language Models from Lexical to Document Simplification." pith.science (2026). https://pith.science/paper/3WPOBEU5

@misc{pith2026250208281,
  author       = {Pith},
  title        = {Pith review of: Redefining Simplicity: Benchmarking Large Language Models from Lexical to Document Simplification},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/3WPOBEU5}},
  note         = {Machine review of arXiv:2502.08281}
}
read the original abstract

Text simplification (TS) refers to the process of reducing the complexity of a text while retaining its original meaning and key information. Existing work only shows that large language models (LLMs) have outperformed supervised non-LLM-based methods on sentence simplification. This study offers the first comprehensive analysis of LLM performance across four TS tasks: lexical, syntactic, sentence, and document simplification. We compare lightweight, closed-source and open-source LLMs against traditional non-LLM methods using automatic metrics and human evaluations. Our experiments reveal that LLMs not only outperform non-LLM approaches in all four tasks but also often generate outputs that exceed the quality of existing human-annotated references. Finally, we present some future directions of TS in the era of LLMs.

Figures

Figures reproduced from arXiv: 2502.08281 by the authors.

Figure 1
Figure 1. Prompt template for LexS. Dataset Instances Substitution per target Min Max Avg English 386 2 22 10.55 Spanish 381 2 19 10.28 Portuguese 386 1 16 8.10 [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Prompt template for SenS. MAP@1. (3) Results The results of all methods are displayed in [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Prompt template for SynS. 2.3 Syntactic Simplification (SynS) (1) Baselines Prior to LLMs, the work of SynS can be divided into two categories. The first one is modeled as a sequence-to￾sequence task where systems are trained on parallel corpora synthesized from knowledge graphs [Narayan et al., 2017], mined from Wikipedia [Botha et al., 2018] and crowd￾sourced [Gao et al., 2021]. The second one relies on a larger s… view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Prompt template for document simplification. [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 3 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. APIO: Automatic Prompt Induction and Optimization for Grammatical Error Correction and Text Simplification

    cs.CL 2025-08 conditional novelty 6.0 of 10

    APIO automatically induces and optimizes instruction-list prompts for grammatical error correction and text simplification, reporting improved scores over prior prompt-based methods on BEA-2019 and ASSET.

  2. A Human-in-the-Loop Corpus for LLM-Based Simplification of Scientific Summaries

    cs.CL 2026-07 accept novelty 5.0 of 10

    A human-in-the-loop corpus of scientific-summary simplifications with original, GPT-simplified, reader-annotated, and expert-edited versions for training and benchmarking simplification systems.

  3. Example-Guided Prompting for Document-Level Text Simplification

    cs.CL 2026-08 conditional novelty 4.0 of 10

    Adding retrieved document-simplification examples to LLM prompts improves SARI scores for four of five tested LLMs, but gains depend heavily on the model.

Reference graph

Works this paper leans on

53 extracted references · 47 canonical work pages · cited by 3 Pith papers

  1. [1]

    Ucca: A semantics-based grammatical annotation scheme

    [Abend and Rappoport, 2013 ] Omri Abend and Ari Rap- poport. Ucca: A semantics-based grammatical annotation scheme. In Proceedings of the 10th International Con- ference on Computational Semantics (IWCS 2013)–Long Papers, pages 1–12,

    work page 2013

  2. [5]

    Lexi: A tool for adaptive, personal- ized text simplification

    [Bingel et al., 2018] Joachim Bingel, Gustavo Paetzold, and Anders Søgaard. Lexi: A tool for adaptive, personal- ized text simplification. In Proceedings of the 27th Inter- national Conference on Computational Linguistics , pages 245–258,

    work page 2018

  3. [7]

    Text simplification tools for spanish

    [Bott et al., 2012] Stefan Bott, Horacio Saggion, and Simon Mille. Text simplification tools for spanish. In LREC, pages 1665–1671,

    work page 2012

  4. [11]

    Paragraph-level sim- plification of medical texts

    [Devaraj et al., 2021] Ashwin Devaraj, Byron C Wallace, Iain J Marshall, and Junyi Jessy Li. Paragraph-level sim- plification of medical texts. In Proceedings of the con- ference. Association for Computational Linguistics. Nort h American Chapter . Meeting , volume 2021, page 4972,

    work page 2021

  5. [12]

    Collabora- tive document simplification using multi-agent systems

    [Fang et al., 2025] Dengzhao Fang, Jipeng Qiang, Xiaoye Ouyang, Yi Zhu, Y unhao Y uan, and Y un Li. Collabora- tive document simplification using multi-agent systems. In Proceedings of the 31st International Conference on Com- putational Linguistics, pages 897–912, Abu Dhabi, UAE, January

    work page 2025

  6. [14]

    Abcd: A graph framework to con- vert complex sentences to a covering set of simple sen- tences

    [Gao et al., 2021] Y anjun Gao, Ting Hao Huang, and Re- becca J Passonneau. Abcd: A graph framework to con- vert complex sentences to a covering set of simple sen- tences. In Joint Conference of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing, ACL-IJCNLP 20...

    work page 2021

  7. [15]

    Text simplification for legal domain: {i} nsights and challenges

    [Garimella et al., 2022] Aparna Garimella, Abhilasha Sancheti, Vinay Aggarwal, Ananya Ganesh, Niyati Chhaya, and Nanda Kambhatla. Text simplification for legal domain: {i} nsights and challenges. In Proceedings of the Natural Legal Language Processing W orkshop 2022, pages 296–304,

    work page 2022

  8. [16]

    Jades: New text simplification dataset in japanese targeted at non-native speakers

    [Hayakawa et al., 2022] Akio Hayakawa, Tomoyuki Kaji- wara, Hiroki Ouchi, and Taro Watanabe. Jades: New text simplification dataset in japanese targeted at non-native speakers. In Proceedings of the W orkshop on T ext Simplifi- cation, Accessibility, and Readability (TSAR-2022) , pages 179–187,

    work page 2022

Show all 53 references
  1. [17]

    Deberta: Decoding-enhanced bert with disentangled attention

    [He et al., 2020] Pengcheng He, Xiaodong Liu, Jianfeng Gao, and Weizhu Chen. Deberta: Decoding-enhanced bert with disentangled attention. arXiv preprint arXiv:2006.03654,

    2020 arXiv

  2. [19]

    Bless: Bench- marking large language models on sentence simplification

    [Kew et al., 2023] Tannon Kew, Alison Chi, Laura V´ asquez- Rodr´ ıguez, Sweta Agrawal, Dennis Aumiller, Fernando Alva-Manchego, and Matthew Shardlow. Bless: Bench- marking large language models on sentence simplification. arXiv preprint arXiv:2310.15773 ,

    2023 arXiv

  3. [20]

    Derivation of new readabil- ity formulas (automated readability index, fog count and flesch reading ease formula) for navy enlisted personnel

    [Kincaid, 1975] JP Kincaid. Derivation of new readabil- ity formulas (automated readability index, fog count and flesch reading ease formula) for navy enlisted personnel. Chief of Naval T echnical Training,

    1975

  4. [22]

    [Laban et al., 2021] Philippe Laban, Tobias Schnabel, Paul Bennett, and Marti A. Hearst. Keep it simple: Unsu- pervised simplification of multi-paragraph text. In Pro- ceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th Internationa...

    2021

  5. [23]

    Joty, Caiming Xiong, and Chien- Sheng Wu

    [Laban et al., 2023] Philippe Laban, Jesse Vig, Wojciech Kryscinski, Shafiq R. Joty, Caiming Xiong, and Chien- Sheng Wu. Swipe: A dataset for document-level simplifi- cation of wikipedia pages. In Annual Meeting of the Asso- ciation for Computational Linguistics ,

    2023

  6. [24]

    BART: Denoising sequence-to-sequence pre-training for natural language generation, translation, and comprehension

    [Lewis et al., 2020] Mike Lewis, Yinhan Liu, Naman Goyal, Marjan Ghazvininejad, Abdelrahman Mohamed, Omer Levy, V eselin Stoyanov, and Luke Zettlemoyer. BART: Denoising sequence-to-sequence pre-training for natural language generation, translation, and comprehension. In Procee...

    2020

  7. [25]

    Mantis at tsar-2022 shared task: Im- proved unsupervised lexical simplification with pretraine d encoders

    [Li et al., 2022] Xiaofei Li, Daniel Wiechmann, Y u Qiao, and Elma Kerz. Mantis at tsar-2022 shared task: Im- proved unsupervised lexical simplification with pretraine d encoders. arXiv preprint arXiv:2212.09855 ,

    2022 arXiv

  8. [26]

    Multilingual lexical sim- plification via paraphrase generation

    [Liu et al., 2023] Kang Liu, Jipeng Qiang, Y un Li, Y unhao Y uan, Yi Zhu, and Kaixun Hua. Multilingual lexical sim- plification via paraphrase generation. In ECAI 2023 - 26th European Conference on Artificial Intelligence , volume 372 of Frontiers in Artificial Intelligence and ...

    2023

  9. [27]

    An unsupervised method for building sen- tence simplification corpora in multiple languages

    [Lu et al., ] Xinyu Lu, Jipeng Qiang, Y un Li, Y unhao Y uan, and Yi Zhu. An unsupervised method for building sen- tence simplification corpora in multiple languages. In Findings of the Association for Computational Linguistics: EMNLP 2021, pages 227–237. [Lu, 2010] Xiaofei Lu....

    2021

  10. [29]

    MUSS: Multilingual unsupervised sentence simplification by min- ing paraphrases

    [Martin et al., 2022] Louis Martin, Angela Fan, ´Eric de la Clergerie, Antoine Bordes, and Benoˆ ıt Sagot. MUSS: Multilingual unsupervised sentence simplification by min- ing paraphrases. In Proceedings of the Thirteenth Lan- guage Resources and Evaluation Conference, pages 1651– 1664,

    2022

  11. [30]

    A document-level text simplifica- tion dataset for japanese

    [Nagai et al., 2024] Y oshinari Nagai, Teruaki Oka, and Mamoru Komachi. A document-level text simplifica- tion dataset for japanese. In Proceedings of the 2024 Joint International Conference on Computational Lin- guistics, Language Resources and Evaluation (LREC- COLING 2024), ...

    2024

  12. [31]

    Cohen, and Anastasia Shimorina

    [Narayan et al., 2017] Shashi Narayan, Claire Gardent, Shay B. Cohen, and Anastasia Shimorina. Split and rephrase. In Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing , pages 606–616, Copenhagen, Denmark, September

    2017

  13. [33]

    Beyond reference: Evaluating high quality translations better than human references

    [Noh et al., 2024] Keonwoong Noh, Seokjin Oh, and Woohwan Jung. Beyond reference: Evaluating high quality translations better than human references. In Pro- ceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 5111–5127,

    2024

  14. [34]

    Automatic text simplification for french: model fine-tuning for simplicity assessment and simpler text generation

    [Ormaechea and Tsourakis, 2024 ] Luc´ ıa Ormaechea and Nikos Tsourakis. Automatic text simplification for french: model fine-tuning for simplicity assessment and simpler text generation. International Journal of Speech T echnology, 27(4):957–976,

    2024

  15. [35]

    Unsupervised lexical simplification for non-nativ e speakers

    [Paetzold and Specia, 2016 ] Gustavo Paetzold and Lucia Specia. Unsupervised lexical simplification for non-nativ e speakers. In Proceedings of the AAAI Conference on Arti- ficial Intelligence , number 1,

    2016

  16. [38]

    Chinese lexical substitution: Dataset and method

    [Qiang et al., 2023] Jipeng Qiang, Kang Liu, Ying Li, Y un Li, Yi Zhu, Y un-Hao Y uan, Xiaocheng Hu, and Xiaoye Ouyang. Chinese lexical substitution: Dataset and method. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing , pages 29–42,

    2023

  17. [39]

    Making it simplext: Implementation and evaluation of a text simplification system for spanish

    [Saggion et al., 2015] Horacio Saggion, Sanja ˇStajner, Ste- fan Bott, Simon Mille, Luz Rello, and Biljana Drndarevic. Making it simplext: Implementation and evaluation of a text simplification system for spanish. ACM Transactions on Accessible Computing (TACCESS), 6(4):1–36,

    2015

  18. [41]

    Improving sequence-to-sequence models for abstractive text summarization using meta heuristic ap- proaches

    [Saxena and Ranjan, 2024 ] Aditya Saxena and Ashutosh Ranjan. Improving sequence-to-sequence models for abstractive text summarization using meta heuristic ap- proaches. arXiv preprint arXiv:2403.16247 ,

    2024 arXiv

  19. [42]

    Exploring german multi-level text simpli- fication

    [Spring et al., 2021] Nicolas Spring, Annette Rios, and Sarah Ebling. Exploring german multi-level text simpli- fication

    2021

  20. [43]

    Simple and effective text simplification using se- mantic and neural methods

    [Sulem et al., 2018] Elior Sulem, Omri Abend, and Ari Rap- poport. Simple and effective text simplification using se- mantic and neural methods. In Proceedings of the 56th Annual Meeting of the Association for Computational Lin- guistics (V olume 1: Long Papers), pages 162–173,

    2018

  21. [44]

    Document-level text simplification: Dataset, crite- ria and baseline

    [Sun et al., 2021] Renliang Sun, Hanqi Jin, and Xiaojun Wan. Document-level text simplification: Dataset, crite- ria and baseline. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing , pages 7997–8013. Association for Computational Linguis- tics,

    2021

  22. [45]

    Problems in current text simplification re- search: New data can help

    [Xu et al., 2015] Wei Xu, Chris Callison-Burch, and Court- ney Napoles. Problems in current text simplification re- search: New data can help. Transactions of the Associa- tion for Computational Linguistics , 3:283–297,

    2015

  23. [46]

    Optimizing sta- tistical machine translation for text simplification

    [Xu et al., 2016] Wei Xu, Courtney Napoles, Ellie Pavlick, Quanze Chen, and Chris Callison-Burch. Optimizing sta- tistical machine translation for text simplification. Trans- actions of the Association for Computational Linguistics , 4:401–415,

    2016

  24. [47]

    Contrastive preference optimization: Pushing the boundaries of llm performance in machine translation

    [Xu et al., 2024] Haoran Xu, Amr Sharaf, Y unmo Chen, Weiting Tan, Lingfeng Shen, Benjamin V an Durme, Ken- ton Murray, and Y oung Jin Kim. Contrastive preference optimization: Pushing the boundaries of llm performance in machine translation. arXiv preprint arXiv:2401.08417 ,

    2024 arXiv

  25. [48]

    Semantic graphs for syntactic simplifi- cation: A revisit from the age of llm

    [Y aoet al., 2024] Peiran Y ao, Kostyantyn Guzhva, and De- nilson Barbosa. Semantic graphs for syntactic simplifi- cation: A revisit from the age of llm. In Proceedings of T extGraphs-17: Graph-based Methods for Natural Lan- guage Processing, pages 105–115,

    2024

  26. [49]

    Bartscore: Evaluating generated text as text generation

    [Y uanet al., 2021] Weizhe Y uan, Graham Neubig, and Pengfei Liu. Bartscore: Evaluating generated text as text generation. ArXiv, abs/2106.11520,

    2021 arXiv

  27. [50]

    Predicting sentence deletions for text simplification using a functional discourse structure

    [Zhang et al., 2022] Bohan Zhang, Prafulla Kumar Choubey, and Ruihong Huang. Predicting sentence deletions for text simplification using a functional discourse structure. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (V olume 2: Shor...

    2022

  28. [51]

    Prompting large language model for machine translation: A case study

    [Zhang et al., 2023] Biao Zhang, Barry Haddow, and Alexandra Birch. Prompting large language model for machine translation: A case study. In International Conference on Machine Learning , pages 41092–41110. PMLR,

    2023

  29. [52]

    Benchmarking large language models for news summarization

    [Zhang et al., 2024] Tianyi Zhang, Faisal Ladhak, Esin Dur- mus, Percy Liang, Kathleen McKeown, and Tatsunori B Hashimoto. Benchmarking large language models for news summarization. Transactions of the Association for Computational Linguistics, 12:39–57,

    2024

  30. [53]

    LSFramework.png

    This figure "LSFramework.png" is available in "png" format from: http://arxiv.org/ps/2502.08281v1

    arXiv

  31. [1975]

    Fine-grained text simplification in french: steps towards a better grammaticality

    [Koptient and Grabar, 2020] Ana¨ ıs Koptient and Natalia Grabar. Fine-grained text simplification in french: steps towards a better grammaticality. In International Sympo- sium on Health Information Management Research ,

    2020

  32. [2002]

    Lsbert: Lexical simpli- fication based on bert

    [Qiang et al., 2021] Jipeng Qiang, Y un Li, Yi Zhu, Y unhao Y uan, Y ang Shi, and Xindong Wu. Lsbert: Lexical simpli- fication based on bert. IEEE/ACM transactions on audio, speech, and language processing, 29:3064–3076,

    2021

  33. [2010]

    Lens: A learnable evaluation met- ric for text simplification

    [Maddela et al., 2023] Mounica Maddela, Y ao Dou, David Heineman, and Wei Xu. Lens: A learnable evaluation met- ric for text simplification. In The 61st Annual Meeting Of The Association F or Computational Linguistics,

    2023

  34. [2012]

    Mcts: A multi- reference chinese text simplification dataset

    [Chong et al., 2024] Ruining Chong, Luming Lu, Liner Y ang, Jinran Nie, Zhenghao Liu, Shuo Wang, Shuhan Zhou, Y aoxin Li, and Erhong Y ang. Mcts: A multi- reference chinese text simplification dataset. In Proceed- ings of the 2024 Joint International Conference on Com- putation...

    2024

  35. [2013]

    Automated text simplification: a survey

    [Al-Thanyyan and Azmi, 2021 ] Suha S Al-Thanyyan and Aqil M Azmi. Automated text simplification: a survey. ACM Computing Surveys (CSUR) , 54(2):1–36,

    2021

  36. [2015]

    Findings of the tsar- 2022 shared task on multilingual lexical simplification

    [Saggion et al., 2023] Horacio Saggion, Sanja ˇStajner, Daniel Ferr´ es, Kim Cheng Sheang, Matthew Shardlow, Kai North, and Marcos Zampieri. Findings of the tsar- 2022 shared task on multilingual lexical simplification. arXiv preprint arXiv:2302.02888 ,

    2023 arXiv

  37. [2016]

    Bleu: a method for au- tomatic evaluation of machine translation

    [Papineni et al., 2002] Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. Bleu: a method for au- tomatic evaluation of machine translation. In Proceedings of the 40th annual meeting of the Association for Compu- tational Linguistics, pages 311–318,

    2002

  38. [2017]

    [Niklaus et al., 2023] Christina Niklaus, Matthias Cetto, Andr´ e Freitas, and Siegfried Handschuh

    Association for Computational Linguistics. [Niklaus et al., 2023] Christina Niklaus, Matthias Cetto, Andr´ e Freitas, and Siegfried Handschuh. Discourse-aware text simplification: From complex sentences to linked propositions. arXiv preprint arXiv:2308.00425 ,

    2023 arXiv

  39. [2018]

    Learning to split and rephrase from wikipedia edit history

    [Botha et al., 2018] Jan A Botha, Manaal Faruqui, John Alex, Jason Baldridge, and Dipanjan Das. Learning to split and rephrase from wikipedia edit history. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 732–737,

    2018

  40. [2020]

    Text simplification of patent documents

    [Kang et al., 2018] Jeongwoo Kang, Achille Souili, and De- nis Cavallucci. Text simplification of patent documents. In Automated Invention for Smart Industries: 18th Inter- national TRIZ Future Conference, TFC 2018, Strasbourg, France, October 29–31, 2018, Proceedings , pages 225–

    2018

  41. [2021]

    UniHD at TSAR-2022 shared task: Is compute all we need for lexical simplification? In TSAR, pages 251– 258,

    [Aumiller and Gertz, 2022 ] Dennis Aumiller and Michael Gertz. UniHD at TSAR-2022 shared task: Is compute all we need for lexical simplification? In TSAR, pages 251– 258,

    2022

  42. [2022]

    Abstract meaning representation for sembanking

    [Banarescu et al., 2013] Laura Banarescu, Claire Bonial, Shu Cai, Madalina Georgescu, Kira Griffitt, Ulf Herm- jakob, Kevin Knight, Philipp Koehn, Martha Palmer, and Nathan Schneider. Abstract meaning representation for sembanking. In Proceedings of the 7th linguistic annota- t...

    2013

  43. [2023]

    Deep non-rigid structure-from-motion: A sequence-to-sequence transla- tion perspective

    [Deng et al., 2024] Hui Deng, Tong Zhang, Y uchao Dai, Ji- awei Shi, Yiran Zhong, and Hongdong Li. Deep non-rigid structure-from-motion: A sequence-to-sequence transla- tion perspective. IEEE transactions on pattern analysis and machine intelligence,

    2024

  44. [2024]

    Document-level planning for text sim- plification

    [Cripwell et al., 2023] Liam Cripwell, Jo¨ el Legrand, and Claire Gardent. Document-level planning for text sim- plification. In Proceedings of the 17th Conference of the European Chapter of the Association for Computational Linguistics, pages 993–1006,

    2023

  45. [2025]

    [Feng et al., 2023] Y utao Feng, Jipeng Qiang, Y un Li, Y un- hao Y uan, and Yi Zhu

    Association for Computational Linguistics. [Feng et al., 2023] Y utao Feng, Jipeng Qiang, Y un Li, Y un- hao Y uan, and Yi Zhu. Sentence simplification via large language models. CoRR, abs/2302.11957,

    2023 arXiv

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.