Pith. sign in

REVIEW 4 major objections 3 minor 45 references

Leveraging Semantic Triples for Private Document Generation with Local Differential Privacy Guarantees

T0 review · 4 major / 3 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read DP-ST claims that privatizing only a document's semantic triples—within a cluster-restricted local-DP neighborhood—yields coherent private text at low privacy budgets.

desk verdict A novel triple-level privatization pipeline with a real privacy-claim problem: the mechanism is neighborhood-DP, not local DP, and the paper needs to own that in the title and abstract. read the letter →

arxiv 2508.20736 v1 pith:YYSJJT6A submitted 2025-08-28 cs.CL

Stephen Meisenbacher , Maulik Chevli , Florian Matthes This is my paper
classification cs.CL
keywords localdifferentialprivacytextprivatizationsemantictriplesExponentialMechanismLLMpost-processingprivacy-utilitytrade-offdocumentgenerationclusteredtriplecorpus
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

DP-ST claims that a document can be privatized under local differential privacy without destroying its usefulness by first distilling it into subject-verb-object (semantic) triples, privatizing each triple with the Exponential Mechanism against a clustered corpus of public triples, and then asking an LLM to rewrite the privatized triples as fluent prose. The central claim is that this divide-and-conquer design achieves coherent, semantically similar outputs at document-level epsilon values far below those previously needed, because the LLM post-processing absorbs much of the noise. The paper also claims that, among the tested local-DP text rewriting methods, DP-ST is the only one that achieves positive privacy-utility trade-offs on average, across five privacy-sensitive datasets and base epsilon values of 0.1, 0.5, and 1. The cost is an explicit relaxation: the DP guarantee applies only within a semantic cluster, called the privatization neighborhood, rather than over the whole text domain.

What carries the argument

Semantic triples, defined here as subject-verb-object extractions such as 'Mark Zuckerberg | founded | Facebook,' are the atomic units of privatization. The method builds a roughly 13.4-million-triple public corpus, embeds the triples, and clusters them; each cluster is a privatization neighborhood. The Exponential Mechanism selects a replacement triple within the cluster based on the cosine similarity between embedded triples. LLM post-processing then fuses the privatized triples into a document, relying on the DP post-processing property.

What would settle it

Run an audit that first records which cluster each input triple is assigned to, then trains a classifier to predict the source cluster from the privatized triple or from the LLM-reconstructed text. If classification accuracy substantially exceeds random guessing at the tested epsilon values, the neighborhood-aware guarantee fails to hide the topic it claims to protect.

Watch

Extended reading notes

Core claim

The central discovery is that restricting the local-DP comparison set to a cluster of semantically related triples makes the privacy-utility trade-off tractable at low budgets. An input document is decomposed into triples; each triple is mapped to the nearest cluster among 50,000–200,000 clusters of public triples; the Exponential Mechanism selects a replacement triple from that cluster using cosine similarity as the utility score; and an LLM reconstructs a coherent text from the privatized triples. Because the LLM sees only privatized triples, post-processing preserves the DP guarantee. The relaxed neighborhood-aware local-DP definition treats only triples in the same cluster as neighbors,

Load-bearing premise

The guarantee only protects a triple from being distinguished from other triples in the same semantic cluster; if the cluster itself is observable, the document's broad topic leaks, and the paper does not formally bound that cluster-level leakage.

Editorial extensions

If this is right

  • If DP-ST is right, local-DP text privatization does not require very large or unbounded privacy budgets; coherent outputs are achievable at base epsilon values of 0.1–1 when the domain is narrowed to a semantic neighborhood.
  • The divide-and-conquer decomposition means privacy protection is focused on the semantic core—who did what to whom—while stylistic attributes and modifiers outside the triples are not covered by the guarantee.
  • LLM reconstruction acts as a free post-processing step: it turns a set of noisy triples into fluent text without increasing the privacy budget, suggesting that similar post-processing could improve other DP text rewriting methods.
  • Cluster size becomes a tunable knob: smaller cluster counts, giving larger neighborhoods, generally improve relative gains, so practitioners can adjust the privacy-utility balance without changing epsilon.
  • The results support the view that coherence, not just lexical or semantic similarity, determines whether DP privatization is worthwhile; methods with high privacy but low coherence yield negative trade-offs.

Reading between the lines

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

  • The neighborhood guarantee hides a triple among semantically similar triples but does not hide the fact that the input belongs to a particular cluster; if cluster centers encode topics, an adversary who recovers the cluster assignment may learn the document's broad topic, and the paper does not formally bound this leakage.
  • Because the public corpus is built from web text and the reconstruction LLM is prompted to fill in gaps, part of the observed privacy may come from semantic abstraction and generic rewriting rather than from the DP mechanism itself; the experiments do not separate these contributions.
  • A testable extension is a cluster-membership audit: train a classifier to predict the source cluster from the privatized triple or reconstructed text, and compare its accuracy to random guessing at each epsilon value.
  • The pipeline is modular, so improved triple extractors or reconstruction models could be swapped in; the same cluster-neighborhood idea could also be applied to other structured representations, such as event tuples or knowledge-graph edges, for private structured data with coherent downstream generation.
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 / 3 minor

Summary. DP-ST is a text privatization pipeline that decomposes input documents into semantic SVO triples, maps each triple to a cluster of public triples via cosine nearest-centroid search, privatizes the triple with an Exponential Mechanism restricted to that cluster, and reconstructs a coherent document from the private triples using an LLM. The authors explicitly introduce a relaxed 'privatization neighborhood' notion: DP indistinguishability holds only between triples assigned to the same cluster. The experimental evaluation compares DP-ST (with three cluster sizes and two LLMs) against TEM, DP-BART, DP-PROMPT, and DP-MLM on five datasets, measuring coherence (G-Eval), cosine similarity, adversarial author/gender inference, and relative gain. The paper reports that DP-ST is the only method with positive average relative gain and that it runs in time comparable to existing methods.

Significance. The engineering contribution is substantial and reproducible: the authors release code for triple extraction, corpus construction, clustering, and the DP-ST method, and the evaluation is broad. If the relaxed neighborhood-based guarantee is accepted as a meaningful formal privacy notion, the divide-and-conquer triple-level design plus LLM post-processing is a promising direction that, unlike the baselines, maintains positive relative gains at the tested document-level budgets. The paper is also transparent in Sections 3.2.1 and 5 that the guarantee is not standard local DP. However, the formal status of the relaxed notion is the central scientific issue: the mechanism is not epsilon-DP over the full input domain, and no analysis quantifies what the deterministic cluster selection leaks. The empirical privacy evaluation measures authorship and gender inference only, not the cluster/topic inference that is most directly threatened by the relaxed guarantee.

major comments (4)
  1. [§3.2.1 and Appendix A] The mechanism is not epsilon-DP over the full input domain. In the standard local DP definition given in Section 2, the inequality must hold for all x,y in the domain. DP-ST selects a cluster deterministically from the input triple and restricts the output support to that cluster. For x in C_i and y in C_j with i≠j, the output supports are disjoint, so for any z at least one of Pr[M(x)=z] and Pr[M(y)=z] is zero and the ratio is 0/0 or infinite. The proof in A.1 considers only x,y in the same cluster. The abstract and Section 1 claim 'local DP guarantees' without this qualification. This is an overclaim unless the relaxed notion is formally defined as a privacy notion with an explicit neighborhood structure and the title/abstract are aligned with it.
  2. [§3.2, §4.3, §5] No bound is provided on leakage from the cluster assignment itself. The cluster is a deterministic function of the input triple, and the output is restricted to that cluster. If clusters are semantically coherent (as intended), an adversary who observes the output can infer that the input lies in a particular topic cluster (e.g., mental health vs. product review). The empirical privacy experiments in Section 4.3 measure authorship and gender inference only; they do not measure cluster or topic inference. The paper should either provide a formal bound on cluster membership inference or clearly state that cluster-level/topic information is not protected and temper the 'strong privacy protections' claim in Section 1.
  3. [Appendix A] The sensitivity reduction from 2 to 1 is not formally established. The text says 'Empirically, we found that the cosine distance between any two vectors in any cluster is always non-negative' and then mentions an explicit check that clips negative values to 0. As written, this is an empirical, not a proof-based, justification. If the utility function is defined as u = max(cos,0), then the range is [0,1] and the l2-sensitivity is 1 by construction. Please state this definition explicitly and remove the empirical claim from the proof.
  4. [Limitations] The fallback path breaks the privacy guarantee for documents from which no semantic triple is extracted. The Limitations section states that such inputs are returned without modification. For these inputs, there is no DP transformation at all. Since some evaluated datasets contain short documents (e.g., Spooky Authors has an average of 30 words per text), this fallback may affect a non-trivial fraction of inputs. The paper should report how often the fallback occurs and explicitly restrict the privacy claim to documents that actually undergo triple decomposition and privatization.
minor comments (3)
  1. [Appendix A] The notation is confusing: P is used both for the set of triples from a sensitive text and for the domain of M: P → V, while the proof sums over z_i ∈ V even though the mechanism's support is the chosen cluster C_i. Please clarify the domain/codomain of the mechanism.
  2. [Table 4] The text says document budgets are computed by 'taking the nearest integer,' but the table reports values such as 57.5, 14.1, 20.8. Please align the description with the actual values or explain the rounding convention.
  3. [Figure 2] The aggregated plots do not show variability, although standard deviations are available in the appendix tables (e.g., the three training runs). Adding error bars or shaded regions would make the comparisons more informative.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: DP-ST's privacy bound follows from a standard Exponential Mechanism application to a disclosed neighborhood relation, and its utility/privacy evaluations are measured against external baselines and benchmarks.

full rationale

The central derivation chain is self-contained. Section 3.2 and Appendix A define the Exponential Mechanism over a fixed cluster of public triples; the epsilon-DP bound is the textbook EM sensitivity argument (cosine utility clipped to [0,1], Delta_u=1), not an equation fitted to the target result. The only relaxation—restricting the neighbor relation to the selected semantic cluster—is explicitly stated in Section 3.2.1 ('we relax this condition and consider only the triples belonging to a cluster as neighbors') and repeated in the Limitations ('we leverage a relaxed neighborhood-aware local DP notion based on privatization neighborhoods'). Since the paper does not claim standard full-domain local DP for cross-cluster inputs, the definitional caveat is disclosed rather than smuggled in. The self-citations (DP-MLM, post-processing work) are used as baselines or prior engineering, not as load-bearing justification for the privacy proof. No fitted parameter is renamed as a prediction. Concerns about whether cluster-level leakage invalidates the practical privacy claim are substantive correctness or interpretability issues, not circularity. Thus there is no circular step requiring a score above 0.

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

The method relies on a public triple corpus and a relaxed privacy notion. No physical or conceptual entity is invented beyond the existing semantic triple representation. The key assumptions are the meaningfulness of the neighborhood-aware guarantee and the empirical sensitivity reduction.

free parameters (3)
  • Number of clusters k = 50000, 100000, 200000
    Selected a priori as privatization parameter; affects cluster size and neighborhood definition.
  • Base epsilon values = 0.1, 0.5, 1
    Used to set document-level budgets scaled by average word count; not fitted to data but chosen for comparison.
  • LSH threshold for triple filtering = 0.4
    Fixed by implementation to deduplicate OpenIE extractions.
assumptions (3)
  • domain assumption Neighborhood-aware DP provides meaningful privacy despite input-dependent cluster selection
    Section 3.2.1 states the relaxation; no formal argument that cluster membership is non-sensitive. The mechanism's output reveals the cluster, allowing different inputs in different clusters to have disjoint output distributions.
  • ad hoc to paper Cosine similarity between any two triples in a cluster is non-negative
    Appendix A claims empirical observation; used to set sensitivity to 1 instead of 2, although clamping of negative values is also implemented.
  • standard math LLM post-processing preserves DP via post-processing property
    Section 3.3 invokes post-processing; this is standard and does not require the LLM to be private.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Leveraging Semantic Triples for Private Document Generation with Local Differential Privacy Guarantees." pith.science (2026). https://pith.science/paper/YYSJJT6A

@misc{pith2026250820736,
  author       = {Pith},
  title        = {Pith review of: Leveraging Semantic Triples for Private Document Generation with Local Differential Privacy Guarantees},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/YYSJJT6A}},
  note         = {Machine review of arXiv:2508.20736}
}
abstract

Many works at the intersection of Differential Privacy (DP) in Natural Language Processing aim to protect privacy by transforming texts under DP guarantees. This can be performed in a variety of ways, from word perturbations to full document rewriting, and most often under local DP. Here, an input text must be made indistinguishable from any other potential text, within some bound governed by the privacy parameter $\varepsilon$. Such a guarantee is quite demanding, and recent works show that privatizing texts under local DP can only be done reasonably under very high $\varepsilon$ values. Addressing this challenge, we introduce DP-ST, which leverages semantic triples for neighborhood-aware private document generation under local DP guarantees. Through the evaluation of our method, we demonstrate the effectiveness of the divide-and-conquer paradigm, particularly when limiting the DP notion (and privacy guarantees) to that of a privatization neighborhood. When combined with LLM post-processing, our method allows for coherent text generation even at lower $\varepsilon$ values, while still balancing privacy and utility. These findings highlight the importance of coherence in achieving balanced privatization outputs at reasonable $\varepsilon$ levels.

Figures

Figures reproduced from arXiv: 2508.20736 by the authors.

Figure 1
Figure 1. An overview of the DP-ST pipeline. In the Preparation stage, publicly available texts are decomposed into semantic triples and stored in a vector database with cluster indices. Then, during Privatization, each extracted triple from an input document is replaced with a differentially private triple achieved through the use of the Exponential Mechanism. All private triples are woven together into a reconstructed priva… view at source ↗
Figure 2
Figure 2. Aggregated experiment results. All figures portray the average scores over our five datasets. Solid lines [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

45 extracted references · 15 canonical work pages

  1. [1]

    Anum Afzal, Ribin Chalumattu, Florian Matthes, and Laura Mascarell. 2024. https://doi.org/10.18653/v1/2024.customnlp4u-1.8 A dapt E val: Evaluating large language models on domain adaptation for text summarization . In Proceedings of the 1st Workshop on Customizable NLP: Progress and Challenges in Customizing NLP for a Domain, Application, Group, or Indiv...

    doi:10.18653/v1/2024.customnlp4u-1.8 2024

  2. [2]

    Gabor Angeli, Melvin Jose Johnson Premkumar, and Christopher D. Manning. 2015. https://doi.org/10.3115/v1/P15-1034 Leveraging linguistic structure for open domain information extraction . In Proceedings of the 53rd Annual Meeting of the Association for Computational Linguistics and the 7th International Joint Conference on Natural Language Processing (Vol...

    doi:10.3115/v1/p15-1034 2015

  3. [3]

    Stefan Arnold, Dilara Yesilbas, and Sven Weinzierl. 2023. https://doi.org/10.18653/v1/2023.trustnlp-1.14 Guiding text-to-text privatization by syntax . In Proceedings of the 3rd Workshop on Trustworthy Natural Language Processing (TrustNLP 2023), pages 151--162, Toronto, Canada. Association for Computational Linguistics

    doi:10.18653/v1/2023.trustnlp-1.14 2023

  4. [4]

    Borja Balle and Yu-Xiang Wang. 2018. https://proceedings.mlr.press/v80/balle18a.html Improving the G aussian mechanism for differential privacy: Analytical calibration and optimal denoising . In Proceedings of the 35th International Conference on Machine Learning, volume 80 of Proceedings of Machine Learning Research, pages 394--403. PMLR

    work page 2018

  5. [5]

    Haohan Bo, Steven H. H. Ding, Benjamin C. M. Fung, and Farkhund Iqbal. 2021. https://doi.org/10.18653/v1/2021.naacl-main.314 ER - AE : Differentially private text generation for authorship anonymization . In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages...

    work page doi:10.18653/v1/2021.naacl-main.314 2021

  6. [6]

    Ricardo Silva Carvalho, Theodore Vasiloudis, Oluwaseyi Feyisetan, and Ke Wang. 2023. https://doi.org/10.1137/1.9781611977653.ch99 TEM : High utility metric differential privacy on text . In Proceedings of the 2023 SIAM International Conference on Data Mining (SDM), pages 883--890. SIAM

    work page doi:10.1137/1.9781611977653.ch99 2023

  7. [7]

    Hyung Won Chung, Le Hou, Shayne Longpre, Barret Zoph, Yi Tay, William Fedus, Eric Li, Xuezhi Wang, Mostafa Dehghani, Siddhartha Brahma, Albert Webson, Shixiang Shane Gu, Zhuyun Dai, Mirac Suzgun, Xinyun Chen, Aakanksha Chowdhery, Sharan Narang, Gaurav Mishra, Adams Yu, and 12 others. 2022. https://doi.org/10.48550/ARXIV.2210.11416 Scaling instruction-fine...

    arXiv doi:10.48550/arxiv.2210.11416 2022

  8. [8]

    Cynthia Dwork. 2006. https://doi.org/10.1007/11787006 Differential privacy . In International colloquium on automata, languages, and programming, pages 1--12. Springer

    doi:10.1007/11787006 2006

Show all 45 references
  1. [9]

    Natasha Fernandes, Mark Dras, and Annabelle McIver. 2019. https://doi.org/10.1007/978-3-030-17138-4_6 Generalised differential privacy for text document processing . In Principles of Security and Trust: 8th International Conference, POST 2019, Held as Part of the European Join...

    2019 doi

  2. [10]

    Oluwaseyi Feyisetan, Borja Balle, Thomas Drake, and Tom Diethe. 2020. https://doi.org/10.1145/3336191.3371856 Privacy- and utility-preserving textual analysis via calibrated multivariate perturbations . In Proceedings of the 13th International Conference on Web Search and Data...

    2020

  3. [11]

    Aristides Gionis, Piotr Indyk, and Rajeev Motwani. 1999. https://www.vldb.org/conf/1999/P49.pdf Similarity search in high dimensions via hashing . In Proceedings of the 25th International Conference on Very Large Data Bases, VLDB '99, page 518–529, San Francisco, CA, USA. Morg...

    1999

  4. [12]

    Aaron Grattafiori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Alex Vaughan, Amy Yang, Angela Fan, Anirudh Goyal, Anthony Hartshorn, Aobo Yang, Archi Mitra, Archie Sravankumar, Artem Korenev, Art...

    2024 arXiv

  5. [13]

    Ivan Habernal. 2021. https://doi.org/10.18653/v1/2021.emnlp-main.114 When differential privacy meets NLP : The devil is in the detail . In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 1522--1528, Online and Punta Cana, Dominican...

    2021 doi

  6. [14]

    Pengcheng He, Xiaodong Liu, Jianfeng Gao, and Weizhu Chen. 2021. https://openreview.net/forum?id=XPZIaotutsD Deberta: Decoding-enhanced bert with disentangled attention . In International Conference on Learning Representations

    2021

  7. [15]

    Dirk Hovy, Anders Johannsen, and Anders S gaard. 2015. https://doi.org/10.1145/2736277.2741141 User review sites as a resource for large-scale sociolinguistic studies . In Proceedings of the 24th International Conference on World Wide Web, WWW '15, page 452–461, Republic and C...

    2015

  8. [16]

    Lijie Hu, Ivan Habernal, Lei Shen, and Di Wang. 2024. https://aclanthology.org/2024.findings-eacl.33 Differentially private natural language models: Recent advances and future directions . In Findings of the Association for Computational Linguistics: EACL 2024, pages 478--499,...

    2024

  9. [17]

    Timour Igamberdiev, Thomas Arnold, and Ivan Habernal. 2022. https://aclanthology.org/2022.coling-1.258/ DP -rewrite: Towards reproducibility and transparency in differentially private text rewriting . In Proceedings of the 29th International Conference on Computational Linguis...

    2022

  10. [18]

    Timour Igamberdiev and Ivan Habernal. 2023. https://doi.org/10.18653/v1/2023.findings-acl.874 DP - BART for privatized text rewriting under local differential privacy . In Findings of the Association for Computational Linguistics: ACL 2023, pages 13914--13934, Toronto, Canada....

    2023 doi

  11. [19]

    Kenny, Shiro Kuriwaki, Cory McCartan, Evan T

    Christopher T. Kenny, Shiro Kuriwaki, Cory McCartan, Evan T. R. Rosenman, Tyler Simko, and Kosuke Imai. 2021. https://doi.org/10.1126/sciadv.abk3283 The use of differential privacy for census data and its impact on redistricting: The case of the 2020 u.s. census . Science Adva...

    2021 doi

  12. [20]

    Oleksandra Klymenko, Stephen Meisenbacher, and Florian Matthes. 2022. https://doi.org/10.18653/v1/2022.privatenlp-1.1 Differential privacy in natural language processing: The story so far . In Proceedings of the Fourth Workshop on Privacy in Natural Language Processing, pages ...

    2022 doi

  13. [21]

    Lewis, Yiming Yang, Tony G

    David D. Lewis, Yiming Yang, Tony G. Rose, and Fan Li. 2004. https://www.jmlr.org/papers/volume5/lewis04a/lewis04a.pdf Rcv1: A new benchmark collection for text categorization research . J. Mach. Learn. Res., 5:361–397

    2004

  14. [22]

    Mike Lewis, Yinhan Liu, Naman Goyal, Marjan Ghazvininejad, Abdelrahman Mohamed, Omer Levy, Veselin Stoyanov, and Luke Zettlemoyer. 2020. https://doi.org/10.18653/v1/2020.acl-main.703 BART : Denoising sequence-to-sequence pre-training for natural language generation, translatio...

    2020 doi

  15. [23]

    Zehan Li, Xin Zhang, Yanzhao Zhang, Dingkun Long, Pengjun Xie, and Meishan Zhang. 2023. https://arxiv.org/abs/2308.03281 Towards general text embeddings with multi-stage contrastive learning . Preprint, arXiv:2308.03281

    2023 arXiv

  16. [24]

    Yang Liu, Dan Iter, Yichong Xu, Shuohang Wang, Ruochen Xu, and Chenguang Zhu. 2023. https://doi.org/10.18653/v1/2023.emnlp-main.153 G -eval: NLG evaluation using gpt-4 with better human alignment . In Proceedings of the 2023 Conference on Empirical Methods in Natural Language ...

    2023 doi

  17. [25]

    Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. 2019. https://arxiv.org/abs/1907.11692 Roberta: A robustly optimized BERT pretraining approach . CoRR, abs/1907.11692

    2019 arXiv

  18. [26]

    Zhi Liu. 2006. https://doi.org/10.24432/C5DS42 Reuter\_50\_50 . UCI Machine Learning Repository

    2006 doi

  19. [27]

    Christopher Manning, Mihai Surdeanu, John Bauer, Jenny Finkel, Steven Bethard, and David McClosky. 2014. https://doi.org/10.3115/v1/P14-5010 The S tanford C ore NLP natural language processing toolkit . In Proceedings of 52nd Annual Meeting of the Association for Computational...

    2014 doi

  20. [28]

    Justus Mattern, Benjamin Weggenmann, and Florian Kerschbaum. 2022. https://doi.org/10.18653/v1/2022.findings-naacl.65 The limits of word level differential privacy . In Findings of the Association for Computational Linguistics: NAACL 2022, pages 867--881, Seattle, United State...

    2022 doi

  21. [29]

    Frank McSherry and Kunal Talwar. 2007. https://doi.org/10.1109/FOCS.2007.66 Mechanism design via differential privacy . In 48th Annual IEEE Symposium on Foundations of Computer Science (FOCS'07), pages 94--103

    2007 doi

  22. [30]

    Stephen Meisenbacher, Maulik Chevli, Juraj Vladika, and Florian Matthes. 2024. https://doi.org/10.18653/v1/2024.findings-acl.554 DP - MLM : Differentially private text rewriting using masked language models . In Findings of the Association for Computational Linguistics: ACL 20...

    2024 doi

  23. [31]

    Stephen Meisenbacher and Florian Matthes. 2024 a . https://doi.org/10.1145/3664476.3669926 Just rewrite it again: A post-processing method for enhanced semantic similarity and privacy preservation of differentially private rewritten text . In Proceedings of the 19th Internatio...

    2024

  24. [32]

    Stephen Meisenbacher and Florian Matthes. 2024 b . https://doi.org/10.18653/v1/2024.emnlp-main.324 Thinking outside of the differential privacy box: A case study in text privatization with language model prompting . In Proceedings of the 2024 Conference on Empirical Methods in...

    2024 doi

  25. [33]

    Guilherme Penedo, Hynek Kydl \' c ek, Loubna Ben allal, Anton Lozhkov, Margaret Mitchell, Colin Raffel, Leandro Von Werra, and Thomas Wolf. 2024. https://openreview.net/forum?id=n6SCkn2QaG The fineweb datasets: Decanting the web for the finest text data at scale . In The Thirt...

    2024

  26. [34]

    Alec Radford, Jeff Wu, Rewon Child, David Luan, Dario Amodei, and Ilya Sutskever. 2019. https://cdn.openai.com/better-language-models/language_models_are_unsupervised_multitask_learners.pdf Language models are unsupervised multitask learners . OpenAI

    2019

  27. [35]

    Nils Reimers and Iryna Gurevych. 2019. https://doi.org/10.18653/v1/D19-1410 Sentence- BERT : Sentence embeddings using S iamese BERT -networks . In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference...

    2019 doi

  28. [36]

    Phillip Schneider, Manuel Klettner, Elena Simperl, and Florian Matthes. 2024. https://aclanthology.org/2024.eacl-short.31/ A comparative analysis of conversational large language models in knowledge-based text generation . In Proceedings of the 18th Conference of the European ...

    2024

  29. [37]

    Hwanjun Song, Hang Su, Igor Shalyminov, Jason Cai, and Saab Mansour. 2024. https://doi.org/10.18653/v1/2024.acl-long.51 F ine S ur E : Fine-grained summarization evaluation using LLM s . In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics...

    2024 doi

  30. [38]

    Samuel Sousa and Roman Kern. 2023. https://doi.org/10.1007/s10462-022-10204-6 How to keep text private? a systematic review of deep learning methods for privacy-preserving natural language processing . Artificial Intelligence Review, 56(2):1427--1492

    2023 doi

  31. [39]

    Saba Sturua, Isabelle Mohr, Mohammad Kalim Akram, Michael Günther, Bo Wang, Markus Krimmel, Feng Wang, Georgios Mastrapas, Andreas Koukounas, Andreas Koukounas, Nan Wang, and Han Xiao. 2024. https://arxiv.org/abs/2409.10173 jina-embeddings-v3: Multilingual embeddings with task...

    2024 arXiv

  32. [40]

    Saiteja Utpala, Sara Hooker, and Pin-Yu Chen. 2023. https://doi.org/10.18653/v1/2023.findings-emnlp.566 Locally differentially private document generation using zero shot prompting . In Findings of the Association for Computational Linguistics: EMNLP 2023, pages 8442--8457, Si...

    2023 doi

  33. [41]

    Doan Nam Long Vu, Timour Igamberdiev, and Ivan Habernal. 2024. https://doi.org/10.18653/v1/2024.findings-emnlp.29 Granularity is crucial when applying differential privacy to text: An investigation for neural machine translation . In Findings of the Association for Computation...

    2024 doi

  34. [42]

    Benjamin Weggenmann, Valentin Rublack, Michael Andrejczuk, Justus Mattern, and Florian Kerschbaum. 2022. https://doi.org/10.1145/3485447.3512232 DP-VAE : Human-readable text anonymization for online reviews with differentially private variational autoencoders . In Proceedings ...

    2022

  35. [43]

    Eric Zhu, Vadim Markovtsev, Aleksey Astafiev, Arham Khan, Chris Ha, Wojciech Łukasiewicz, Adam Foster, Sinusoidal36, Spandan Thakur, Stefano Ortolani, Titusz, Vojtech Letal, Zac Bentley, fpug, hguhlich, long2ice, oisincar, Ron Assa, Senad Ibraimoski, and 8 others. 2024. https:...

    2024 doi

  36. [44]

    online" 'onlinestring :=

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

  37. [45]

    write newline

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

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.