Pith. sign in

REVIEW 5 major objections 4 minor 30 references

Automated Feedback Loops to Protect Text Simplification with Generative AI from Information Loss

T0 review · 5 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read Inserting every named entity that a generative model dropped from a simplified health text restores the most information, according to cosine-similarity and ROUGE-1 comparisons.

desk verdict A buildable but over-interpreted feedback loop: the all-entities insertion win likely reflects token quantity and circular similarity, not robust semantic correctness. read the letter →

arxiv 2505.16172 v1 pith:YKJYVDTK submitted 2025-05-22 cs.CL cs.AI

Abhay Kumara Sri Krishna Nandiraju , Gondy Leroy , David Kauchak , Arif Ahmed This is my paper · ORCID
classification cs.CLcs.AI
keywords textsimplificationgenerativeAIbiomedicalnamedentitiesinformationlossfeedbackloopChatGPTcosinesimilarityROUGE
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 argues that the information loss introduced by generative AI in health-text simplification can be caught and repaired automatically. On 50 rheumatology texts simplified by gpt-4-0613, the authors compare five repair strategies: inserting all missing named entities, all missing words, the top three entities ranked by ChatGPT, three random entities, and k random entities. They report that inserting all missing entities gives the highest document-level cosine similarity (0.9162) and ROUGE-1 score (0.6555), while word-level insertion gives the best summary-level alignment. If this holds, patient-facing simplified health text can be made more faithful to the source with a fully automated loop, reducing the risk that critical medical details are silently dropped.

What carries the argument

The machinery is a three-stage feedback loop. First, gpt-4-0613 simplifies the source text. Second, missing information is identified in two ways: named entities are extracted with scispacy from the original and simplified text, and the set difference yields the missing entity set; missing words are stemmed tokens that occur at least twice in the original text but fewer than twice in the simplified text. Third, gpt-4-0613 regenerates the simplified text with the missing elements inserted, guided by a structured prompt. The comparison is carried by two automated metrics, cosine similarity over 384-dimensional MiniLM embeddings and ROUGE-1 F1, computed at both the full-document and BART-summary level.

What would settle it

A reader-comprehension experiment would settle it: lay participants answer multiple-choice questions about the original, simplified, and A1-repaired texts; if comprehension of A1-repaired text is not significantly higher than comprehension of the simplified text, the claim that A1 better preserves information fails despite its higher automated similarity scores.

Watch

Extended reading notes

Core claim

The paper's central claim is that exhaustive, content-aware reinsertion of missing named entities, called approach A1, is the most effective repair strategy: it restored the highest semantic similarity and unigram overlap with the original full documents, with mean cosine similarity 0.9162 and mean ROUGE-1 F1 0.6555, compared with 0.8471 and 0.5063 before insertion. Adding all missing words, approach A2, was best on summaries, with cosine similarity 0.8058 and ROUGE-1 0.4213. The paper concludes that current entity-recognition tools can identify what was dropped but are not reliable for ranking which dropped entities matter most, since the top-three ranked approach performed no better than inserting three random entities. This supports a repair strategy of inserting the full set of missing named entities rather than a selectively chosen subset.

Load-bearing premise

The load-bearing premise is that higher cosine similarity and ROUGE-1 agreement with the original text means the regenerated text actually preserves the information that matters; no human reading or task-based measure is used to confirm that.

Editorial extensions

If this is right

  • If A1's result holds, exhaustively reinserting missing named entities rather than selecting a few is the right repair design for preserving document-level content.
  • Word-level repair can complement entity-level repair, since A2 produced better summary-level alignment, suggesting fine-grained terms carry information that matters in condensed evaluation.
  • Ranking missing entities with an LLM gives no measurable advantage over random selection when only three entities are added, so importance ranking should not be trusted unless independently validated.
  • A fully automated simplification-plus-repair pipeline is within reach for health texts: detection uses off-the-shelf entity recognition and frequency counts, and regeneration uses prompting.

Reading between the lines

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

  • Because the augmented text is created by inserting tokens drawn from the original, the reported gains in cosine similarity and ROUGE-1 may partly reflect lexical overlap rather than restored understanding; a comprehension test would tell whether the mechanical gains translate into reader benefit.
  • The missing-word rule is a simple frequency threshold (at least two occurrences in the original, fewer than two in the simplified text); varying that threshold could change which words are reinserted and is a natural next test.
  • A hybrid that adds both all missing entities and all missing words is the paper's stated future direction, and the separate results suggest such a hybrid should improve both document-level and summary-level metrics simultaneously.
  • The approach is likely to transfer best to other text genres where domain terminology is dense and reliably recognized by biomedical named-entity recognition; in low-resource domains with poor entity recognition, different detectors may be needed.
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

5 major / 4 minor

Summary. The paper proposes an automated feedback loop for health text simplification: a large language model (gpt-4-0613) simplifies a source text, a detection step identifies named entities and words that are present in the original but missing or rare in the simplified text, and a regeneration step uses gpt-4-0613 to insert the missing elements. Five insertion strategies are compared (A1: all missing entities; A2: all missing words; A3: top-3 ranked entities; A4: three random entities; A5: |W_missing| random entities) on 50 rheumatology articles from the BMJ. Evaluation is by cosine similarity (all-MiniLM-L6-v2 embeddings) and ROUGE-1 against the original text, at both full-document and BART-summary levels. The reported results show A1 achieving the highest document-level cosine similarity (0.9162) and ROUGE-1 (0.6555), while A2 achieves the best summary-level scores (0.8058 and 0.4213). The paper concludes that adding all missing entities is the most effective and robust method for improving the semantic correctness of simplified biomedical texts.

Significance. The problem addressed is important: large language models may delete information that is crucial for understanding health texts, and automated methods to detect and repair such omissions would be practically valuable. The paper's strengths are a clearly specified pipeline, the use of a real biomedical corpus, and the inclusion of random-insertion controls. However, the central evaluation concern raised in the stress-test note lands on reading the manuscript: the missing elements are defined as tokens drawn from the original text, and the evaluation metrics compare the augmented text back to that same original, so the measured advantage of A1 is at least partially mechanical. The paper provides no human or task-based validation of its core assumption that higher similarity to the original reflects more complete information. If that assumption were validated and the count-controlled comparisons reproduced the ranking, the proposed feedback loop would be a useful contribution; as it stands, the evidence supports a more limited, descriptive claim about the metrics rather than the strong conclusion about semantic correctness.

major comments (5)
  1. [Secs. 3.1-3.3, Table 1] The evaluation metric is not independent of the construction of the augmented texts. Missing entities and missing words are defined as elements of the original text that are absent or rare in the simplified text (Sec. 3.1), the augmentation step inserts those exact elements (Sec. 3.2), and cosine similarity and ROUGE-1 are then computed against the original (Sec. 4.2). Since the inserted tokens are copied from the reference, both metrics are inflated by construction. The paper states the assumption that 'a higher similarity to the original text reflects more complete information' (Sec. 3), but no human comprehension study, task-based measure, or expert judgment is provided to validate this assumption. Consequently, the current results do not support the conclusion in Sec. 7 that A1 improves 'semantic correctness'; they show only that A1 increases similarity to the original under these two metrics.
  2. [Sec. 3.2, Table 1] A1 is not compared against a count-matched control. A4 inserts exactly three random entities, A5 inserts |W_missing| random entities, and A1 inserts all |SE_missing| missing entities. The paper does not report the average sizes of SE_missing and W_missing, so it is possible that A1 simply inserts more original-derived material than the other approaches. If so, the higher cosine similarity (0.9162) and ROUGE-1 (0.6555) for A1 could be a quantity effect rather than evidence that entities are more informative than words or that all-entity insertion is superior. A2 versus A5 is approximately count-matched (both use |W_missing| items), which makes that comparison more interpretable, but the A1 comparisons lack the same control.
  3. [Table 1, Secs. 5-7] The summary-level results contradict the 'most effective and robust' conclusion. At the summary level A2 outperforms A1 on both cosine similarity (0.8058 vs. 0.7609) and ROUGE-1 (0.4213 vs. 0.3908), and A1 has the lowest summary-level cosine similarity of all five approaches. The explanation in Sec. 6 that A1 captures broader semantic concepts while A2 captures fine-grained details is not supported by any additional analysis, such as a breakdown of entity types or a comparison of concepts retained. The paper should either reconcile this reversal with explicit evidence or qualify the conclusion that A1 is the best overall method.
  4. [Table 1, Sec. 5] The results are reported as means over N=50 texts without standard deviations, confidence intervals, or significance tests. Adjacent values, such as the document-level cosine similarity of A1 (0.9162) versus A2 (0.8990), or the ROUGE-1 of A4 (0.5454) versus A3 (0.5364), may be within sampling noise. Paired tests (e.g., bootstrap or Wilcoxon signed-rank) and effect sizes are needed before the relative ranking of approaches can be considered reliable.
  5. [Sec. 4.2] The paper labels its metric 'ROUGE-1' but defines it with set overlap: U_overlap = U_T0 ∩ U_TS, where counts are not used. Standard ROUGE-1 (Lin, 2004) is computed over multiset/bag-of-words counts and accounts for term frequency. Because the augmented texts are produced by inserting tokens from the original, a set-based overlap measure can overstate the contribution of a single copied occurrence and thereby inflate the scores of methods that insert more distinct original tokens. The authors should either compute standard ROUGE-1 or call the metric 'unigram-set F1' and discuss the consequences for their comparison.
minor comments (4)
  1. [References] The reference list contains a placeholder entry 'Author, F., Author, S.: Title of a proceedings paper. In: Editor, F., Editor, S. (eds.) CONFERENCE 2016...' appended to reference [1]; this should be removed and the actual reference list completed.
  2. [Sec. 4.2] The notation for ROUGE-1 is inconsistent: the text uses T0 and TS almost interchangeably, and 'R1-r' appears to denote recall in one place but is not clearly distinguished from precision and F1. Defining R1-recall, R1-precision, and R1-F1 with consistently named subscripts would improve readability.
  3. [Sec. 3.1] The threshold for missing words (present twice or more in the original but fewer than twice in the simplified text) is arbitrary, and no sensitivity analysis is provided. At minimum, the authors should acknowledge this as a free parameter and state how sensitive the missing-word set is to it.
  4. [Sec. 4.2, Table 1] Reporting the average numbers of missing entities and missing words per document, along with text lengths, would make the count-matched issue in A1/A5 easier to assess and would help readers interpret the absolute magnitude of the reported differences.

Circularity Check

1 steps flagged · score 6.0 of 10

The document-level 'best' claim for A1 reduces to feeding original-derived tokens into a similarity-to-original metric.

  1. self definitional [Sec. 3.1–3.3 (W_missing and SE_missing definitions; A1/A2 insertion; cosine/ROUGE evaluation) and Sec. 5/7 (best-method claim)]
    "W_missing = {w ∈ W_original | f(w,T_original) ≥ 2 and f(w,T_simplified) < 2} ... SE_missing = SE_original ∖ SE_simplified ... A1: All the missing entities (SE_missing) are added to the simplified text to generate an improved simplified version ... We assume for this study that a higher similarity to the original text reflects more complete information."

    The 'missing' content is defined as the set difference between the original and simplified texts. Approaches A1 and A2 instruct the generator to insert exactly those original-text-derived tokens back into the simplified text. The evaluation metrics then compare the augmented text with the original: ROUGE-1 counts unigram overlap with the original, and cosine similarity measures embedding closeness to the original. Any inserted token is, by definition, a unigram of the original, so it directly increases the overlap and pulls the embedding toward the original. The observed improvement, especially A1's document-level gain, is therefore a mechanical consequence of the definitions rather than an independent measurement of information value or semantic correctness.

full rationale

The paper is not globally circular: the pipeline for identifying missing entities and words, the regeneration prompts, and the comparison of insertion policies are concrete and implementable, and the self-citations to prior simplification work are contextual rather than load-bearing. No uniqueness theorem or author-imported ansatz carries the argument. The circularity is localized to the evaluation of the central claim. Missing content is defined as original-minus-simplified content; the augmentation inserts that same original content; and the metrics reward overlap and similarity with the original. Thus A1's document-level superiority is in substantial part a restatement of the construction, not evidence that entity insertion produces better semantic correctness. The absence of a count-matched entity control (A4 uses three entities, A5 uses |W_missing|, A1 uses all |SE_missing|) means the comparison cannot separate the type of inserted content from the quantity of inserted original-derived tokens. The summary-level reversal, where A2 outperforms A1, further shows that the conclusion depends on which representation of the original is used. A task-based or human evaluation of information completeness would be needed to support the 'semantic correctness' conclusion. Score 6 reflects partial circularity: the central claim's document-level evidence reduces to the construction, while the regeneration step still involves a generative model and the controls provide some comparative signal.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The central comparison rests on two hand-set thresholds (word frequency threshold 2, number of ranked entities 3) and a control-matching design choice, plus four domain assumptions: similarity-to-original tracks information completeness; set-difference identifies all critical omissions; BART summaries are faithful; and regenerated text introduces no new errors. No invented entities are introduced.

free parameters (3)
  • frequency threshold for missing words = 2
    W_missing = {w : f(w,T_original) >= 2 and f(w,T_simplified) < 2} (Sec. 3.1); chosen without tuning or sensitivity analysis.
  • number of ranked entities inserted (A3) = 3
    Top-3 entities by gpt-4-0613 importance ranking are inserted (Sec. 3.2); the count is arbitrary and not varied.
  • number of random entities inserted (A5 control) = |W_missing|
    A5 inserts k random entities where k = |W_missing| to match A2's word count (Sec. 3.2); this matching choice is a design decision that shapes the comparison.
assumptions (4)
  • domain assumption Higher similarity to the original text reflects more complete information (and better simplification).
    Stated explicitly in Sec. 3 before the identification methods; this is the bedrock of all evaluation, yet similarity metrics reward lexical overlap and can penalize legitimate simplification.
  • domain assumption The set-difference of scispacy entities and the frequency threshold w >= 2 in original and w < 2 in simplified capture all crucial missing information.
    Sec. 3.1 defines W_missing and SE_missing; no validation that omitted entities or words are those that matter for comprehension, nor that the threshold is appropriate.
  • domain assumption BART-generated summaries faithfully condense the core meaning of the original and augmented texts, so summary-level metrics are meaningful.
    Sec. 3.3 uses BART to create summaries for two-step evaluation; no evaluation of summary quality for this dataset is provided.
  • domain assumption GPT-4 reinsertion produces fluent text without introducing factual errors or hallucinations.
    The regeneration in Sec. 3.2 is not checked for hallucinations or grammatical quality; prior work cited by the paper (refs 24, 26, 28) notes LLMs can hallucinate.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Automated Feedback Loops to Protect Text Simplification with Generative AI from Information Loss." pith.science (2026). https://pith.science/paper/YKJYVDTK

@misc{pith2026250516172,
  author       = {Pith},
  title        = {Pith review of: Automated Feedback Loops to Protect Text Simplification with Generative AI from Information Loss},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/YKJYVDTK}},
  note         = {Machine review of arXiv:2505.16172}
}
read the original abstract

Understanding health information is essential in achieving and maintaining a healthy life. We focus on simplifying health information for better understanding. With the availability of generative AI, the simplification process has become efficient and of reasonable quality, however, the algorithms remove information that may be crucial for comprehension. In this study, we compare generative AI to detect missing information in simplified text, evaluate its importance, and fix the text with the missing information. We collected 50 health information texts and simplified them using gpt-4-0613. We compare five approaches to identify missing elements and regenerate the text by inserting the missing elements. These five approaches involve adding missing entities and missing words in various ways: 1) adding all the missing entities, 2) adding all missing words, 3) adding the top-3 entities ranked by gpt-4-0613, and 4, 5) serving as controls for comparison, adding randomly chosen entities. We use cosine similarity and ROUGE scores to evaluate the semantic similarity and content overlap between the original, simplified, and reconstructed simplified text. We do this for both summaries and full text. Overall, we find that adding missing entities improves the text. Adding all the missing entities resulted in better text regeneration, which was better than adding the top-ranked entities or words, or random words. Current tools can identify these entities, but are not valuable in ranking them.

Figures

Figures reproduced from arXiv: 2505.16172 by the authors.

Figure 1
Figure 1. The figure shows a structured prompt guiding the augmentation of simplified text with missing entities or words while retaining original context and linguistic simplicity We evaluate five different approaches(A1 to A5) to insert the missing information back into the text to generate an improved simplified text that has better semantic similarity and content overlap with the original text. In each of the five approac… view at source ↗
Figure 2
Figure 2. The figure shows a prompt guiding the evaluation and selection of the most semantically valuable entities to enhance the simplified text. Approaches 4 and 5 are two control conditions in which we insert random information. These are necessary to verify that adding any information does not result in increased similarity. 1) A4: three random entities are chosen randomly from the set of missing entities (𝑆𝐸𝑚𝑖𝑠𝑠𝑖𝑛𝑔) and… view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

30 extracted references · 24 canonical work pages

  1. [1]

    Feuerriegel, S., Hartmann, J., Janiesch, C., & Zschech, P. (2024). Generative ai. Business & Information Systems Engineering, 66(1), 111 -126.Author, F., Author, S.: Title of a proceedings paper. In: Editor, F., Editor, S. (eds.) CONFERENCE 2016, LNCS, vol. 9999, pp. 1–13. Springer, Heidelberg (2016)

    work page 2024

  2. [2]

    Watanangura, P., Vanichrudee, S., Minteer, O., Sringamdee, T., Thanngam, N., & Siriborvornratanakul, T. (2023). A comparative survey of text summarization techniques. SN Computer Science, 5(1), 47

    work page 2023

  3. [3]

    & Weston, J

    Roller, S., Dinan, E., Goyal, N., Ju, D., Williamson, M., Liu, Y., ... & Weston, J. (2020). Recipes for building an open-domain chatbot. arXiv preprint arXiv:2004.13637

    arXiv 2020

  4. [4]

    Clark, E., Ji, Y., & Smith, N. A. (2018, June). Neural te xt generation in stories using entity representations as context. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long Papers) (pp. 2250-2260)

    work page 2018

  5. [5]

    Edmundson, H.P. (1969). New Methods in Automatic Extracting. J. ACM, 16, 264-285

    work page 1969

  6. [6]

    Luhn, H. P. (1958). The automatic creation of literature abstracts. IBM Journal of research and development, 2(2), 159-165

    work page 1958

  7. [7]

    Flesch reading ease and the Flesch Kincaid grade level

    Readable. Flesch reading ease and the Flesch Kincaid grade level. . 2021 [cited 2023 April 29]; Available from: https://readable.com/readability/flesch-reading-ease-flesch-kincaid- grade-level/

    work page 2021

  8. [8]

    The Technique of Clear Writing

    Gunning, Robert (1952). The Technique of Clear Writing. McGraw-Hill. pp. 36–37

    work page 1952

Show all 30 references
  1. [9]

    A Formula for Predicting Readability

    Dale E; Chall J (1948). "A Formula for Predicting Readability". Educational Research Bulletin. 27: 11–20+28

    1948

  2. [10]

    A., & Barai, P

    Ahmed, A., Leroy, G., Lee, S., Harber, P., Kauchak, D., Rains, S. A., & Barai, P. (20 24). Influence of Audio Speech Rate and Source Text Difficulty on Health Information Comprehension and Retention

  3. [11]

    A., Harber, P., Kauchak, D., & Barai, P

    Ahmed, A., Leroy, G., Rains, S. A., Harber, P., Kauchak, D., & Barai, P. (2024). Effects of Added Emphasis and Pause in Audio Delivery of Health Information. AMIA Summits on Translational Science Proceedings, 2024, 54

    2024

  4. [12]

    Y., Kauchak, D., Stone, J., Harber, P.,

    Ahmed, A., Leroy, G., Lu, H. Y., Kauchak, D., Stone, J., Harber, P., ... & Chitr oda, B. (2023). Audio delivery of health information: An NLP study of information difficulty and bias in listeners. Procedia computer science, 219, 1509-1517

    2023

  5. [13]

    E., Kauchak, D., Mouradi, O., & Just, M

    Leroy, G., Endicott, J. E., Kauchak, D., Mouradi, O., & Just, M. (2013). User evaluation of the e ffects of a text simplification algorithm using term familiarity on perception, understanding, learning, and information retention. Journal of medical Internet research , 15(7), e2569

    2013

  6. [14]

    Sulem, E., Abend, O., & Rappoport, A. (2018). Semantic structural evalua tion for text simplification. arXiv preprint arXiv:1810.05022. Automated Feedback Loops to Protect Text 13

    2018 arXiv

  7. [15]

    Leroy, G., & Endicott, J. E. (2012, January). Combining NLP with evidence-based methods to find text metrics related to perceived and actual text difficulty. In Proceedings of the 2nd ACM SIGHIT International Health Informatics Symposium (pp. 749-754)

    2012

  8. [16]

    Kauchak, D., Leroy, G., & Hogue, A. (2017). Measuring text difficulty using parse‐tree frequency. Journal of the Association for Information Science and Technology, 68(9), 2088- 2100

    2017

  9. [17]

    Leroy, G., Ka uchak, D., Harber, P., Pal, A., & Shukla, A. (2024). Text and Audio Simplification: Human vs. ChatGPT. AMIA Summits on Translational Science Proceedings, 2024, 295

    2024

  10. [18]

    (2014, January)

    Kauchak, D., Mouradi, O., Pentoney, C., & Leroy, G. (2014, January). Text simplification tools: Using machine learning to discover features that identify difficult text. In 2014 47th Hawaii international conference on system sciences (pp. 2616-2625). IEEE

    2014

  11. [19]

    Mukherjee, P., Leroy, G., & Kauchak, D. (2018). Using lexical chains to identify text difficulty: a corpus statistics and classification study. IEEE journal of biomedical and health informatics, 23(5), 2164-2173

    2018

  12. [20]

    G., Soroush, A., Elias, P

    Tang, L., Sun, Z., Idnay, B., Nestor, J. G., Soroush, A., Elias, P. A., ... & Peng, Y. (2023). Evaluating large language models on me dical evidence summarization. NPJ digital medicine, 6(1), 158

    2023

  13. [21]

    Z., Rasheed, U., Bian, J.,

    Tariq, R., Malik, S., Roy, M., Islam, M. Z., Rasheed, U., Bian, J., ... & Zhang, R. (2023, June). Assessing chatgpt for text summarization, simplification and extraction tasks. In 2023 IEEE 11th International Conference on Healthcare Informatics (ICHI) (pp. 746-749). IEEE

    2023

  14. [22]

    Radford, A., Narasimhan, K., Salimans, T., & Sutskever, I. (2018). Improving language understanding by generative pre-training

    2018

  15. [23]

    D., Dhariwal, P.,

    Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J. D., Dhariwal, P., ... & Amodei, D. (2020). Language models are few-shot learners. Advances in neural information processing systems, 33, 1877-1901

    2020

  16. [24]

    & Ayre, J

    Stanceski, K., Zhong, S., Zhang, X., Khadra, S., Tracy, M., Koria, L., ... & Ayre, J. (2024). The q uality and safety of using generative AI to produce patient -centred discharge instructions. npj Digital Medicine, 7(1), 329

    2024

  17. [25]

    M., Baskharoun, S., Zhao, Y., Austrian, J., Aphinyanaphongs, Y.,

    Zaretsky, J., Kim, J. M., Baskharoun, S., Zhao, Y., Austrian, J., Aphinyanaphongs, Y., ... & Feldman, J. (2024). Generative artific ial intelligence to transform inpatient discharge summaries to patient -friendly language and format. JAMA network open , 7(3), e240357- e240357

    2024

  18. [26]

    R., Dobbs, T

    Ali, S. R., Dobbs, T. D., Hutchings, H. A., & Whitaker, I. S. (2023). Using ChatGPT to write patient clinic letters. The Lancet Digital Health, 5(4), e179-e181

    2023

  19. [27]

    Feng, Y., Qiang, J., Li, Y., Yuan, Y., & Zhu, Y. (2023). Sentence simplification via large language models. arXiv preprint arXiv:2302.11957

    2023 arXiv

  20. [28]

    & Liu , T

    Huang, L., Yu, W., Ma, W., Zhong, W., Feng, Z., Wang, H., ... & Liu , T. (2025). A survey on hallucination in large language models: Principles, taxonomy, challenges, and open questions. ACM Transactions on Information Systems, 43(2), 1-55

    2025

  21. [29]

    Lin, C. Y. (2004, July). Rouge: A package for automatic evaluation of summaries. In Text summarization branches out (pp. 74-81)

    2004

  22. [30]

    Lewis, M., Liu, Y., Goyal, N., Ghazvininejad, M., Mohamed, A., Levy, O., & Zettlemoyer, L. (2019). Bart: Denoising sequence -to-sequence pre -training for natural language generation, translation, and comprehension. arXiv preprint arXiv:1910.13461

    2019 arXiv

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.