Pith. sign in

REVIEW 5 major objections 5 minor 2 cited by

Fake News Detection After LLM Laundering: Measurement and Explanation

T0 review · 5 major / 5 minor · reviewed 2026-08-10 · deepseek-v4-flash

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

Pith's one-line read The paper claims that LLM paraphrasing makes fake news harder to detect, with sentiment shifts as the likely reason.

desk verdict Systematic but under-supported: the key laundering result is carried by undocumented COVID-19 experiments with suspicious table values. read the letter →

arxiv 2501.18649 v1 pith:DP7AQ7T7 submitted 2025-01-29 cs.CL cs.AIcs.IRcs.LG

Rupak Kumar Das , Jonathan Dodge This is my paper · ORCID
classification cs.CLcs.AIcs.IRcs.LG
keywords fakenewsdetectionLLMlaunderingparaphraseevasionsentimentshiftBERTScoreLIMEexplanationsmisinformationtextclassification
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

LLM paraphrasing—rewriting fake news with GPT, Llama, or Pegasus—makes it harder for fake news detectors to label correctly than human-written text. On a COVID-19 misinformation dataset, all 17 detectors lose F1 on every paraphrased version, with Pegasus-paraphrased text hardest to detect. In a second, six-label political dataset, Pegasus still produces the lowest F1 for most detectors, though the pattern is less uniform. The paper's proposed mechanism, supported by LIME explanations, is that paraphrases often shift sentiment while BERTScore still reports high semantic similarity, so detectors no longer see the negative cues that marked the text as fake. The authors release the paraphrased datasets, the sentiment-shift metric, and threshold heuristics for others to build on.

What carries the argument

The load-bearing object is the sentiment-shift score $S = (P_h - N_h) - (P_l - N_l)$ from Equation (1), where $P$ and $N$ are the positive and negative probabilities returned by a sentiment analyzer for human text ($h$) and LLM paraphrase ($l$). The paper uses it as a difference-of-differences: $|S| > 1$ is taken to mean a flipped sentiment and $|S| > 0.5$ as probably flipped. Around a pair of detections, LIME local explanations supply the qualitative link from shifted sentiment words to misclassification, and BERTScore's F1 supplies the semantic-similarity side that makes the 'high similarity but shifted sentiment' combination apparent.

What would settle it

Take the paraphrases the S metric flags as flipped, have human annotators label their sentiment, and compare detector predictions on flipped versus non-flipped paraphrases with BERTScore held high; if S-flagged flips are often not actually flipped, or detectors misclassify high-BERTScore paraphrases with no sentiment flip, the paper's mechanism is falsified.

Watch

Extended reading notes

Core claim

The discovery the paper argues for is that LLM laundering is an effective evasion attack: rewriting fake news with an LLM paraphrase step degrades detector performance relative to human-written text. On the COVID-19 dataset the effect is uniform: every detector's F1 drops for GPT, Llama, and Pegasus paraphrases, and Pegasus causes the largest drops. On the LIAR dataset the ordering is less consistent, but Pegasus gives the worst F1 for 11 of 17 detectors. The paper further claims that GPT produces the most semantically similar paraphrases by BERTScore, and that LIME explanations point to sentiment shift as the reason detectors fail: words carrying negative sentiment in fake news are replaced by positive or neutral wording in paraphrases, while BERTScore still reports high similarity. The paper presents the sentiment-shift metric S as a first pass and explicitly notes it is not validated.

Load-bearing premise

The explanation for detection failure rests on the unvalidated assumption that the sentiment analyzer's scores and the $|S| > 0.5$ and $|S| > 1$ thresholds correctly identify when a paraphrase flips sentiment, and that this flip is what caused the detector to change its prediction.

Editorial extensions

If this is right

  • Defenders who build fake news pipelines should evaluate detectors on paraphrased test sets, since paraphrasing is a cheap evasion step.
  • Paraphrase quality metrics that ignore sentiment will overestimate how well a paraphrase preserves the label-relevant content of fake news.
  • Detectors that lean on negative sentiment cues are specifically vulnerable to LLM paraphrases that shift sentiment.
  • Pegasus-style abstractive paraphrasing poses a higher current evasion risk than GPT or Llama in the measured settings.
  • The released paraphrased datasets let other researchers measure detector robustness without repeating the paraphrase step.

Reading between the lines

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

  • An untested extension is to add a sentiment-similarity term to BERTScore or a similar metric; if sentiment shift is the failure mechanism, such a metric should predict detector errors better than BERTScore alone.
  • If the mechanism is causal, detectors trained on sentiment-preserving or sentiment-controlled paraphrases should recover much of the lost F1; the paper does not run that experiment.
  • The LIAR results suggest the evasion effect is domain-dependent—short political statements behave differently from longer news text—so conclusions about laundering should be re-checked per text genre.
  • Calibrating the S thresholds against human sentiment judgments would turn the proposed explanation into a testable measurement; the paper leaves that step for 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

5 major / 5 minor

Summary. The paper investigates whether paraphrasing fake news texts with LLMs (GPT, Llama, Pegasus), a process the authors call 'LLM laundering', degrades the ability of fake news detectors to classify them correctly. The authors evaluate 17 detectors on two datasets (COVID-19 misinformation and LIAR), compare human-written versus paraphrased texts on accuracy, F1, precision, and recall, measure paraphrase quality with BERTScore, and use LIME explanations plus a proposed sentiment-shift metric to suggest why detection fails. They report that paraphrasing generally reduces detector F1, that Pegasus paraphrases are hardest to detect, that GPT produces the highest BERTScore paraphrases, and that sentiment shifts paired with high BERTScore are a possible explanation. They also release augmented datasets.

Significance. If the headline result holds, the paper has practical value for misinformation defense by quantifying a concrete attack surface and by showing that current semantic-similarity metrics can miss sentiment flip. The study is also useful as a broad comparison of detectors and paraphrasers. The authors are transparent about some limitations (Section 7) and release the paraphrased datasets, which supports reproducibility. However, the strength of the central claims is currently limited by unresolved evaluation issues, missing statistical support, and anecdotal evidence for the sentiment-shift mechanism.

major comments (5)
  1. [Section 3.2 and Table 1] The paper does not document a train/test split for the COVID-19 dataset; the only mention of a pre-split is for LIAR ('The creators have pre-split the dataset into train, test, and validation'). If detectors were trained and evaluated on the same human-written texts, the very high human-text F1 scores in Table 1 (e.g., BERT .930) could reflect memorization, and the lower scores on paraphrased texts would be an out-of-distribution artifact rather than evidence that LLM laundering evades detection. The authors must state the exact split for COVID-19 and report held-out results. Additionally, in every row of Table 1, Accuracy, F1, Precision, and Recall are identical to three decimals; this is highly unusual and suggests a possible evaluation bug, so the authors should clarify the metric computation or provide confusion matrices.
  2. [Tables 1 and 2] All results are from single runs with no error bars, confidence intervals, or significance tests. In LIAR, accuracies around .25 are close to the majority-class baseline (the most frequent class is about 20.7% of the data), so differences such as BERT F1 .232 on human text versus .251 on GPT-paraphrased text may be within noise. The authors should provide multiple seeds or bootstrap intervals and include a majority-class baseline in every table to calibrate the reader's assessment of the reported differences.
  3. [Section 4.5 and Figure 5] The sentiment-shift explanation is based on two hand-selected examples, one from each dataset. Two instances cannot support the general claim that sentiment shift is a reason for detection failure, and Section 7 concedes that 'a more comprehensive study is needed to ascertain that claim.' The abstract and conclusion nevertheless present sentiment shift as a contribution; the authors should either provide systematic quantitative evidence or reframe it explicitly as an untested hypothesis.
  4. [Equation (1) and Section 4.5] The S metric is unvalidated, as admitted in Section 7, and the thresholds |S| > 1 'definitely' and |S| > 0.5 'more probable than not' are ad hoc with no calibration evidence. The reported percentages (0.45% and 6.86%) are computed only for the Human-GPT condition shown in Figure 6, and the statement that 'all other configurations have patterns' is not substantiated with figures or statistics. This weakens the claimed link between sentiment shift and paraphrase quality.
  5. [Section 6, compared with Section 4.1 and Table 2] The conclusion that 'paraphrasing tends to decrease classification accuracy' is not supported by the LIAR results, where several detectors show higher F1 on paraphrased text (e.g., BERT .232 human versus .251 GPT; T5 .236 human versus .262 Llama). In fact, Section 4.1 states that 'no source was easier or harder to detect consistently' for LIAR. The conclusion needs to be qualified by dataset and paraphraser, or the authors should aggregate results across datasets in a way that accounts for these inconsistencies.
minor comments (5)
  1. [Section 3.1] The paraphrase generation details are missing: the specific GPT and Llama model versions, prompts, decoding parameters (temperature, top-p, max length), and the number of paraphrases generated are not reported, which makes replication difficult even with the released datasets.
  2. [Figure 6] Only the Human-GPT condition is plotted for the sentiment-shift analysis; if other paraphraser conditions are said to follow the same pattern, the authors should include those plots or provide aggregated statistics.
  3. [Section 3.2] The CNN architecture description is unclear: 'The input layer consisted of 1024 units with ReLU activation' suggests a dense layer rather than a convolutional layer; please specify the actual architecture, including kernel sizes and number of filters.
  4. [Section 7] The phrase '(.27% Accuracy)' is ambiguous; presumably the intended meaning is 27% accuracy, not 0.27%.
  5. [Equation (1)] The notation in Equation (1) is inconsistent: the text says 'each of which can come from human (e.g., Nh) or LLM (e.g., Pl)', but the formula uses Ph, Nh, Pl, Nl; please define all subscripts explicitly.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: all central claims are direct empirical measurements with no fitted-input or self-citation reduction.

full rationale

I walked the claimed derivation chain and found no step where a prediction or first-principles result reduces to its own inputs. The paper's central claim—that LLM paraphrasing degrades detector F1 relative to human text—is a direct measurement on two public datasets, with no fitted parameter being relabeled as a prediction. The sentiment-shift explanation (Section 4.5, Equation 1) is generated from LIME observations and then tested with an independent HuggingFace sentiment analyzer; the S metric and its thresholds are admittedly unvalidated (Section 7), but an unvalidated or ad hoc auxiliary metric is a validity limitation, not a circular derivation. No load-bearing self-citation appears: the only reused artifact is a T5 classifier adopted from an external GitHub repository (Section 3.2). The suspicious numeric coincidences in Table 1 and the absent COVID-19 train/test-split documentation are correctness or reproducibility concerns, not circularity. Therefore the paper earns no circularity points.

Assumptions & free parameters 1 free parameters · 4 assumptions · 0 invented entities

The paper introduces one ad hoc numeric rule (the S thresholds) and relies on several unstated assumptions about its sentiment analyzer, BERTScore, and the LIAR testbed. No new physical or conceptual entities are introduced.

free parameters (1)
  • sentiment flip thresholds = |S| > 1 and |S| > 0.5
    Chosen by hand in Section 4.5 to classify 'definitely flipped' and 'more probable than not' sentiment flips; no derivation or validation, and the paper admits the S metric is unvalidated.
assumptions (4)
  • domain assumption The sentiment analyzer (distilbert-based, ref [69]) produces scores that faithfully represent the sentiment of news text and paraphrases.
    Invoked in Section 4.5 to compute S and label flips; no validation on this domain.
  • domain assumption BERTScore F1 is an adequate measure of semantic similarity for paraphrase quality; the paper uses it as ground truth in RQ4.
    The paper relies on ref [21] and then undercuts this assumption in Section 5.2, which is internally consistent but still an assumption.
  • domain assumption The detectors are trained on original human-written text and tested on paraphrased text without leakage; the train/test split for COVID is not stated.
    The methodology section (Section 3) describes data preprocessing but does not describe a train/test split for the COVID-19 dataset.
  • domain assumption LIAR's six-label setup, despite near-random baseline, is a valid environment for comparing detector robustness.
    Section 4.1 and Limitations call LIAR 'rather strange' yet the paper uses its F1 differences to rank models in RQ2/RQ3.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Fake News Detection After LLM Laundering: Measurement and Explanation." pith.science (2026). https://pith.science/paper/DP7AQ7T7

@misc{pith2026250118649,
  author       = {Pith},
  title        = {Pith review of: Fake News Detection After LLM Laundering: Measurement and Explanation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/DP7AQ7T7}},
  note         = {Machine review of arXiv:2501.18649}
}
read the original abstract

With their advanced capabilities, Large Language Models (LLMs) can generate highly convincing and contextually relevant fake news, which can contribute to disseminating misinformation. Though there is much research on fake news detection for human-written text, the field of detecting LLM-generated fake news is still under-explored. This research measures the efficacy of detectors in identifying LLM-paraphrased fake news, in particular, determining whether adding a paraphrase step in the detection pipeline helps or impedes detection. This study contributes: (1) Detectors struggle to detect LLM-paraphrased fake news more than human-written text, (2) We find which models excel at which tasks (evading detection, paraphrasing to evade detection, and paraphrasing for semantic similarity). (3) Via LIME explanations, we discovered a possible reason for detection failures: sentiment shift. (4) We discover a worrisome trend for paraphrase quality measurement: samples that exhibit sentiment shift despite a high BERTSCORE. (5) We provide a pair of datasets augmenting existing datasets with paraphrase outputs and scores. The dataset is available on GitHub

Figures

Figures reproduced from arXiv: 2501.18649 by the authors.

Figure 1
Figure 1. Methodology to assess the efficacy of fake news detectors [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. (Top): Performance of fake news detec￾tors on human-written and LLM-paraphrased text on COVID-19 dataset. (Bottom): Same, but on LIAR dataset For the COVID-19 dataset, LLM-based detectors (BERT, Llama, and T5) and deep learning-based models (LSTM and CNN) all perform well ( [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Distribution of FBERT score for all para￾phrasers on COVID-19 dataset. Higher is better. 0.60 0.65 0.70 0.75 0.80 0.85 0.90 0.95 1.00 GPT 0 1000 Frequency 0.60 0.65 0.70 0.75 0.80 0.85 0.90 0.95 1.00 Llama 0 1000 Frequency 0.60 0.65 0.70 0.75 0.80 0.85 0.90 0.95 1.00 Pegasus 0 1000 Frequency [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Distribution of FBERT score for all para￾phrasers on LIAR dataset 15 of the 17 detectors exhibit the second lowest F1 score for the Llama-paraphrased text for the COVID￾19 data set. On the other hand, GPT-paraphrased fake news is easy to detect for most of the classifi…
Figure 5
Figure 5. Figure 5: (Top Left): LIME output of the BERT model on human-written news (Top Right): LIME output of the BERT model on Llama-paraphrased news (Bottom Left): LIME output of the LSTM model on human-written news (Bottom Right): LIME output of the LSTM model on GPT-paraphrased news…
Figure 6
Figure 6. Figure 6: FBERT score vs sentiment shift (Human￾GPT) on COVID-19 dataset. Here, we plotted only this configuration, as all other configurations have patterns. lation. Assuming a three tuple (+, 0, -), we compute a difference of differences between positive senti￾ment P and negat…

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 2 Pith papers

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

  1. Large Language Models in Misinformation Ecosystems: Misuse, Defense, and Vulnerability

    cs.CR 2026-07 conditional novelty 5.0 of 10

    A role-layer survey unifies LLM misuse, LLM-based defense, and LLM-centric verification vulnerabilities across content, social, evidence, and workflow layers, then lists three open challenges.

  2. Debate-to-Detect: Reformulating Misinformation Detection as a Real-World Debate with Large Language Models

    cs.CL 2025-05 conditional novelty 5.0 of 10

    A structured multi-agent debate framework with domain-specialized AI agents and a five-dimension scoring rubric improves LLM-based fake news detection by several F1 points.

Reference graph

Works this paper leans on

69 extracted references · 52 canonical work pages · cited by 2 Pith papers

  1. [1]

    Joint copying and restricted gener- ation for paraphrase

    Ziqiang Cao, Chuwei Luo, Wenjie Li, and Su- jian Li. Joint copying and restricted gener- ation for paraphrase. In Proceedings of the AAAI Conference on Artificial Intelligence, vol- ume 31, 2017

    work page 2017

  2. [2]

    Semantic parsing via paraphrasing

    Jonathan Berant and Percy Liang. Semantic parsing via paraphrasing. In Proceedings of the 52nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1415–1425, 2014

    work page 2014

  3. [3]

    Open question answering over curated and extracted knowledge bases

    Anthony Fader, Luke Zettlemoyer, and Oren Etzioni. Open question answering over curated and extracted knowledge bases. In Proceed- ings of the 20th ACM SIGKDD international conference on Knowledge discovery and data mining, pages 1156–1165, 2014

    work page 2014

  4. [4]

    Answering questions with complex semantic constraints on open knowledge bases

    Pengcheng Yin, Nan Duan, Ben Kao, Junwei Bao, and Ming Zhou. Answering questions with complex semantic constraints on open knowledge bases. In Proceedings of the 24th ACM international on conference on informa- tion and knowledge management, pages 1301– 1310, 2015

    work page 2015

  5. [5]

    Qanet: Combin- ing local convolution with global self-attention for reading comprehension

    Adams Wei Yu, David Dohan, Minh-Thang Luong, Rui Zhao, Kai Chen, Mohammad Norouzi, and Quoc V Le. Qanet: Combin- ing local convolution with global self-attention for reading comprehension. arXiv preprint arXiv:1804.09541, 2018

    arXiv 2018

  6. [6]

    Adversarial example gener- ation with syntactically controlled paraphrase networks

    Mohit Iyyer, John Wieting, Kevin Gimpel, and Luke Zettlemoyer. Adversarial example gener- ation with syntactically controlled paraphrase networks. arXiv preprint arXiv:1804.06059 , 2018

    arXiv 2018

  7. [7]

    Natural language pro- cessing: state of the art, current trends and challenges

    Diksha Khurana, Aditya Koli, Kiran Khatter, and Sukhdev Singh. Natural language pro- cessing: state of the art, current trends and challenges. Multimedia tools and applications, 82(3):3713–3744, 2023

    work page 2023

  8. [8]

    Openai report details election interference efforts, hoaxes

    Ina Fried. Openai report details election interference efforts, hoaxes. https://www.axios.com/2024/10/09/ openai-election-interference-political-misinformation , 2024

    work page 2024

Show all 69 references
  1. [9]

    Language models are few-shot learners

    Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neu- ral information processing systems, 33:1877– 1901, 2020

    1901

  2. [10]

    Bert: Pre- training of deep bidirectional transformers for language understanding

    Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre- training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018

    2018 arXiv

  3. [11]

    Exploring the limits of transfer learning with a unified text-to-text transformer

    Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of ma- chine learning research, 21(140):1–67, 2020

    2020

  4. [12]

    Llama: Open and efficient foundation language models

    Hugo Touvron, Thibaut Lavril, Gautier Izac- ard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023

    2023 arXiv

  5. [13]

    Defending against neural fake news

    Rowan Zellers, Ari Holtzman, Hannah Rashkin, Yonatan Bisk, Ali Farhadi, Franziska Roesner, and Yejin Choi. Defending against neural fake news. Advances in neural informa- tion processing systems, 32, 2019

    2019

  6. [14]

    Fak- ing fake news for real fake news detection: Propaganda-loaded training data generation

    Kung-Hsiang Huang, Kathleen McKeown, Preslav Nakov, Yejin Choi, and Heng Ji. Fak- ing fake news for real fake news detection: Propaganda-loaded training data generation. arXiv preprint arXiv:2203.05386, 2022

    2022 arXiv

  7. [15]

    Can llm-generated misinformation be detected? arXiv preprint arXiv:2309.13788, 2023

    Canyu Chen and Kai Shu. Can llm-generated misinformation be detected? arXiv preprint arXiv:2309.13788, 2023

    2023 arXiv

  8. [16]

    Fake news detectors are biased against texts generated by large language models

    Jinyan Su, Terry Yue Zhuo, Jonibek Mansurov, Di Wang, and Preslav Nakov. Fake news detectors are biased against texts generated by large language models. arXiv preprint arXiv:2309.08674, 2023

    2023 arXiv

  9. [17]

    The efficacy of detecting ai- generated fake news using transfer learning

    Jake Stewart, Nikita Lyubashenko, and George Stefanek. The efficacy of detecting ai- generated fake news using transfer learning. Issues in Information Systems, 24(2), 2023

    2023

  10. [18]

    Disinformation detection: An evolv- ing challenge in the age of llms

    Bohan Jiang, Zhen Tan, Ayushi Nirmal, and Huan Liu. Disinformation detection: An evolv- ing challenge in the age of llms. arXiv preprint arXiv:2309.15847, 2023

    2023 arXiv

  11. [19]

    Bleu: a method for auto- matic evaluation of machine translation

    Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. Bleu: a method for auto- matic evaluation of machine translation. InPro- ceedings of the 40th annual meeting of the As- sociation for Computational Linguistics, pages 311–318, 2002

    2002

  12. [20]

    Rouge: A package for auto- matic evaluation of summaries

    Chin-Yew Lin. Rouge: A package for auto- matic evaluation of summaries. In Text summa- rization branches out, pages 74–81, 2004

    2004

  13. [21]

    Bertscore: Evaluating text generation with bert

    Tianyi Zhang, Varsha Kishore, Felix Wu, Kil- ian Q Weinberger, and Yoav Artzi. Bertscore: Evaluating text generation with bert. arXiv preprint arXiv:1904.09675, 2019

    1904 arXiv

  14. [22]

    A study of translation edit rate with targeted hu- man annotation

    Matthew Snover, Bonnie Dorr, Richard Schwartz, Linnea Micciulla, and John Makhoul. A study of translation edit rate with targeted hu- man annotation. In Proceedings of the 7th Con- ference of the Association for Machine Transla- tion in the Americas: Technical Papers, pages 22...

    2006

  15. [23]

    A survey on evaluation metrics for machine translation

    Seungjun Lee, Jungseob Lee, Hyeonseok Moon, Chanjun Park, Jaehyung Seo, Sugyeong Eo, Seonmin Koo, and Heuiseok Lim. A survey on evaluation metrics for machine translation. Mathematics, 11(4):1006, 2023

    2023

  16. [24]

    Meant 2.0: Accurate semantic mt evaluation for any output language

    Chi-kiu Lo. Meant 2.0: Accurate semantic mt evaluation for any output language. In Pro- ceedings of the second conference on machine translation, pages 589–597, 2017

    2017

  17. [25]

    Paraphrasing questions using given and new information

    Kathleen McKeown. Paraphrasing questions using given and new information. American Journal of Computational Linguistics, 9(1):1– 10, 1983

    1983

  18. [26]

    Unt: Sub- finder: Combining knowledge sources for au- tomatic lexical substitution

    Samer Hassan, Andras Csomai, Carmen Banea, Ravi Sinha, and Rada Mihalcea. Unt: Sub- finder: Combining knowledge sources for au- tomatic lexical substitution. In Proceedings of the fourth international workshop on semantic evaluations (SemEval-2007), pages 410–413, 2007

    2007

  19. [27]

    Paraphrase generation with deep reinforcement learning

    Zichao Li, Xin Jiang, Lifeng Shang, and Hang Li. Paraphrase generation with deep reinforcement learning. arXiv preprint arXiv:1711.00279, 2017

    2017 arXiv

  20. [28]

    A deep generative framework for paraphrase generation

    Ankush Gupta, Arvind Agarwal, Prawaan Singh, and Piyush Rai. A deep generative framework for paraphrase generation. In Pro- ceedings of the aaai conference on artificial intelligence, volume 32, 2018

    2018

  21. [29]

    Un- supervised paraphrase generation using pre- trained language models

    Chaitra Hegde and Shrikumar Patil. Un- supervised paraphrase generation using pre- trained language models. arXiv preprint arXiv:2006.05477, 2020

    2006 arXiv

  22. [30]

    Neural paraphrase generation with stacked residual lstm networks

    Aaditya Prakash, Sadid A Hasan, Kathy Lee, Vivek Datla, Ashequl Qadir, Joey Liu, and Oladimeji Farri. Neural paraphrase generation with stacked residual lstm networks. arXiv preprint arXiv:1610.03098, 2016

    2016 arXiv

  23. [31]

    Deep residual learning for image recognition

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE con- ference on computer vision and pattern recog- nition, pages 770–778, 2016

    2016

  24. [32]

    Paraphrase generation with latent bag of words

    Yao Fu, Yansong Feng, and John P Cunning- ham. Paraphrase generation with latent bag of words. Advances in Neural Information Pro- cessing Systems, 32, 2019

    2019

  25. [33]

    Neural syntac- tic preordering for controlled paraphrase gener- ation

    Tanya Goyal and Greg Durrett. Neural syntac- tic preordering for controlled paraphrase gener- ation. arXiv preprint arXiv:2005.02013, 2020

    2005 arXiv

  26. [34]

    Decomposable neural paraphrase genera- tion

    Zichao Li, Xin Jiang, Lifeng Shang, and Qun Liu. Decomposable neural paraphrase genera- tion. arXiv preprint arXiv:1906.09741, 2019

    1906 arXiv

  27. [35]

    Generate rather than retrieve: Large language models are strong context generators

    Wenhao Yu, Dan Iter, Shuohang Wang, Yi- chong Xu, Mingxuan Ju, Soumya Sanyal, Chenguang Zhu, Michael Zeng, and Meng Jiang. Generate rather than retrieve: Large language models are strong context generators. arXiv preprint arXiv:2209.10063, 2022

    2022 arXiv

  28. [36]

    Para- phrasing with large language models

    Sam Witteveen and Martin Andrews. Para- phrasing with large language models. arXiv preprint arXiv:1911.09661, 2019

    1911 arXiv

  29. [37]

    How large language models are transforming machine-paraphrased plagiarism

    Jan Philip Wahle, Terry Ruas, Frederic Kirstein, and Bela Gipp. How large language models are transforming machine-paraphrased plagiarism. arXiv preprint arXiv:2210.03568, 2022

    2022 arXiv

  30. [38]

    Pag-llm: Paraphrase and aggregate with large language models for minimizing intent clas- sification errors

    Vikas Yadav, Zheng Tang, and Vijay Srinivasan. Pag-llm: Paraphrase and aggregate with large language models for minimizing intent clas- sification errors. In Proceedings of the 47th International ACM SIGIR Conference on Re- search and Development in Information Re- trieval, pa...

    2024

  31. [39]

    Tinybert: Distilling bert for natu- ral language understanding

    Xiaoqi Jiao, Yichun Yin, Lifeng Shang, Xin Jiang, Xiao Chen, Linlin Li, Fang Wang, and Qun Liu. Tinybert: Distilling bert for natu- ral language understanding. arXiv preprint arXiv:1909.10351, 2019

    1909 arXiv

  32. [40]

    Investigating paraphrasing-based data augmentation for task- oriented dialogue systems

    Liane V ogel and Lucie Flek. Investigating paraphrasing-based data augmentation for task- oriented dialogue systems. In International Conference on Text, Speech, and Dialogue , pages 476–488. Springer, 2022

    2022

  33. [41]

    Submodular optimization-based diverse paraphrasing and its effectiveness in data augmentation

    Ashutosh Kumar, Satwik Bhattamishra, Manik Bhandari, and Partha Talukdar. Submodular optimization-based diverse paraphrasing and its effectiveness in data augmentation. In Pro- ceedings of the 2019 Conference of the North American Chapter of the Association for Com- putational...

    2019

  34. [42]

    Improving data augmentation for low resource speech-to-text translation with diverse para- phrasing

    Chenggang Mi, Lei Xie, and Yanning Zhang. Improving data augmentation for low resource speech-to-text translation with diverse para- phrasing. Neural Networks , 148:194–205, 2022

    2022

  35. [43]

    T5w: A paraphras- ing approach to oversampling for imbalanced text classification

    Annapurna P Patil, Shreekant Jere, Reshma Ram, and Shruthi Srinarasi. T5w: A paraphras- ing approach to oversampling for imbalanced text classification. In 2022 IEEE International Conference on Electronics, Computing and Communication Technologies (CONECCT) , pages 1–6. IEEE, 2022

    2022

  36. [44]

    Paranmt-50m: Pushing the limits of paraphrastic sentence em- beddings with millions of machine translations

    John Wieting and Kevin Gimpel. Paranmt-50m: Pushing the limits of paraphrastic sentence em- beddings with millions of machine translations. arXiv preprint arXiv:1711.05732, 2017

    2017 arXiv

  37. [45]

    Automatically constructing a corpus of sentential paraphrases

    Bill Dolan and Chris Brockett. Automatically constructing a corpus of sentential paraphrases. In Third international workshop on paraphras- ing (IWP2005), 2005

    2005

  38. [46]

    Ppdb: The paraphrase database

    Juri Ganitkevitch, Benjamin Van Durme, and Chris Callison-Burch. Ppdb: The paraphrase database. In Proceedings of the 2013 confer- ence of the north american chapter of the asso- ciation for computational linguistics: Human language technologies, pages 758–764, 2013

    2013

  39. [47]

    Parasci: A large scientific paraphrase dataset for longer paraphrase generation

    Qingxiu Dong, Xiaojun Wan, and Yue Cao. Parasci: A large scientific paraphrase dataset for longer paraphrase generation. arXiv preprint arXiv:2101.08382, 2021

    2021 arXiv

  40. [48]

    Gathering and generating paraphrases from twitter with application to normalization

    Wei Xu, Alan Ritter, and Ralph Grishman. Gathering and generating paraphrases from twitter with application to normalization. In Proceedings of the sixth workshop on building and using comparable corpora, pages 121–128, 2013

    2013

  41. [49]

    Cgmh: Constrained sentence generation by metropolis-hastings sampling

    Ning Miao, Hao Zhou, Lili Mou, Rui Yan, and Lei Li. Cgmh: Constrained sentence generation by metropolis-hastings sampling. In Proceed- ings of the AAAI Conference on Artificial Intel- ligence, volume 33, pages 6834–6842, 2019

    2019

  42. [50]

    Euca: The end-user-centered explainable ai framework

    Weina Jin, Jianyu Fan, Diane Gromala, Philippe Pasquier, and Ghassan Hamarneh. Euca: The end-user-centered explainable ai framework. arXiv preprint arXiv:2102.02437, 2021

    2021 arXiv

  43. [51]

    why should i trust you?

    Marco Tulio Ribeiro, Sameer Singh, and Carlos Guestrin. " why should i trust you?" explaining the predictions of any classifier. In Proceed- ings of the 22nd ACM SIGKDD international conference on knowledge discovery and data mining, pages 1135–1144, 2016

    2016

  44. [52]

    New explainability method for bert- based model in fake news detection

    M Szczepanski, M Pawlicki, R Kozik, and M Choras. New explainability method for bert- based model in fake news detection. sci. rep. 11 (1), 23705 (2021)

    2021

  45. [53]

    A unified approach to in- terpreting model predictions

    Scott Lundberg. A unified approach to in- terpreting model predictions. arXiv preprint arXiv:1705.07874, 2017

    2017 arXiv

  46. [54]

    Combat covid-19 infodemic using explainable natural language processing models

    Jackie Ayoub, X Jessie Yang, and Feng Zhou. Combat covid-19 infodemic using explainable natural language processing models. Informa- tion Processing & Management, 58(4):102569, 2021

    2021

  47. [55]

    Explainable machine learning for fake news detection

    Julio CS Reis, André Correia, Fabrício Mu- rai, Adriano Veloso, and Fabrício Benevenuto. Explainable machine learning for fake news detection. In Proceedings of the 10th ACM conference on web science, pages 17–26, 2019

    2019

  48. [56]

    Axiomatic attribution for deep networks

    Mukund Sundararajan, Ankur Taly, and Qiqi Yan. Axiomatic attribution for deep networks. In International conference on machine learn- ing, pages 3319–3328. PMLR, 2017

    2017

  49. [57]

    Learning important features through propagating activation differences

    Avanti Shrikumar, Peyton Greenside, and An- shul Kundaje. Learning important features through propagating activation differences. In International conference on machine learning, pages 3145–3153. PMlR, 2017

    2017

  50. [58]

    A causal framework for explaining the predic- tions of black-box sequence-to-sequence mod- els

    David Alvarez-Melis and Tommi S Jaakkola. A causal framework for explaining the predic- tions of black-box sequence-to-sequence mod- els. arXiv preprint arXiv:1707.01943, 2017

    2017 arXiv

  51. [59]

    A song of (dis) agree- ment: Evaluating the evaluation of explainable artificial intelligence in natural language pro- cessing

    Michael Neely, Stefan F Schouten, Maurits Bleeker, and Ana Lucic. A song of (dis) agree- ment: Evaluating the evaluation of explainable artificial intelligence in natural language pro- cessing. In HHAI2022: Augmenting Human Intellect, pages 60–78. IOS Press, 2022

    2022

  52. [60]

    Interpreting recurrent and attention-based neural models: a case study on natural language inference

    Reza Ghaeini, Xiaoli Z Fern, and Prasad Tade- palli. Interpreting recurrent and attention-based neural models: a case study on natural language inference. arXiv preprint arXiv:1808.03894, 2018

    2018 arXiv

  53. [61]

    Model explainability in deep learning based nat- ural language processing

    Shafie Gholizadeh and Nengfeng Zhou. Model explainability in deep learning based nat- ural language processing. arXiv preprint arXiv:2106.07410, 2021

    2021 arXiv

  54. [62]

    Fighting an infodemic: Covid-19 fake news dataset, 2020

    Parth Patwa, Shivam Sharma, Srinivas PYKL, Vineeth Guptha, Gitanjali Kumari, Md Shad Akhtar, Asif Ekbal, Amitava Das, and Tanmoy Chakraborty. Fighting an infodemic: Covid-19 fake news dataset, 2020

    2020

  55. [63]

    liar, liar pants on fire

    William Yang Wang. “liar, liar pants on fire”: A new benchmark dataset for fake news detection. arXiv preprint arXiv:1705.00648, 2017

    2017 arXiv

  56. [64]

    Nltk: The natu- ral language toolkit

    Edward Loper and Steven Bird. Nltk: The natu- ral language toolkit. arXiv preprint cs/0205028, 2002

    2002 arXiv

  57. [65]

    A survey on text classification algorithms: From text to predictions

    Andrea Gasparetto, Matteo Marcuzzo, Alessan- dro Zangari, and Andrea Albarelli. A survey on text classification algorithms: From text to predictions. Information, 13(2):83, 2022

    2022

  58. [66]

    Pegasus: Pre-training with ex- tracted gap-sentences for abstractive summa- rization

    Jingqing Zhang, Yao Zhao, Mohammad Saleh, and Peter Liu. Pegasus: Pre-training with ex- tracted gap-sentences for abstractive summa- rization. In International conference on ma- chine learning, pages 11328–11339. PMLR, 2020

    2020

  59. [67]

    Scikit-learn: Machine learning in python

    Fabian Pedregosa, Gaël Varoquaux, Alexan- dre Gramfort, Vincent Michel, Bertrand Thirion, Olivier Grisel, Mathieu Blondel, Pe- ter Prettenhofer, Ron Weiss, Vincent Dubourg, et al. Scikit-learn: Machine learning in python. Journal of machine learning research, 12(Oct):2825–2830, 2011

    2011

  60. [68]

    flan- t5-base-imdb-text-classification

    Mohammad Taghizadeh. flan- t5-base-imdb-text-classification. https://github.com/M-Taghizadeh/ flan-t5-base-imdb-text-classification , 2023

    2023

  61. [69]

    distilbert-base-multilingual- cased-sentiments-student (revision 2e33845), 2023

    Lik Xun Yuan. distilbert-base-multilingual- cased-sentiments-student (revision 2e33845), 2023. URL: https://huggingface. co/lxyuan/distilbert-base-multilingual-cased- sentiments-student. doi, 10

    2023

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.