Pith. sign in

REVIEW 3 major objections 21 references

Lexical and structural fingerprints stay competitive for near-duplicates, while frozen BGE hashes better preserve ranked similarity under rewriting—at higher cost.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · grok-4.5

2026-07-10 08:21 UTC pith:QHNC7I4A

load-bearing objection Useful packaged IR benchmark with real scorer diagnostics; the rewriting half of the trade-off is thinner than the abstract implies. the 3 major comments →

arxiv 2607.08382 v1 pith:QHNC7I4A submitted 2026-07-09 cs.IR

H3D: Benchmarking Unsupervised Text Hashing for Fine-Grained Document Deduplication

Qianren Mao , Jiaxun Lyu , Junnan Liu , Zhijun Chen , Jingzheng Li , Hanwen Hao , Bo Li This is my paper
classification cs.IR
keywords text hashingdocument deduplicationunsupervised hashingMinHashSimHashBGE embeddingsscientific documentsranking evaluation
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

H3D is a unified unsupervised text-hashing benchmark for fine-grained scientific document deduplication. It compares rule-based fingerprints (MinHash, SimHash, Winnowing, FuzzyHash, FlyHash) with codes from frozen BGE embeddings quantized by deterministic binarization or random-hyperplane LSH, all under the same query–candidate ranking protocol on CSFCube and RELISH. Ranking quality (MAP, NDCG@20), runtime, and progressive text compression are reported together. The central finding is a consistent trade-off: surface-oriented hashes work well when wording and structure largely match, while semantic-sensitive codes keep similarity better when abstracts are rewritten or shortened, paying more compute. The work also shows that similarity scorers are not interchangeable and can become rank-equivalent only for particular hash representations, so method and scorer must be judged jointly.

Core claim

Under one shared query–candidate protocol on CSFCube and RELISH, unsupervised non-learning lexical and structural fingerprints remain competitive for near-duplicate scientific matching, whereas BGE-BIHash and BGE-LSHash better preserve similarity under content rewriting and controlled compression, at higher computational cost. Different similarity measures can be rank-equivalent for some hash outputs and divergent for others, so fair comparison requires reporting generator and scorer together.

What carries the argument

H3D’s unified query–candidate ranking pipeline: each method produces a fingerprint (fixed-rule or frozen BGE then quantized), a type-specific scorer ranks the candidate pool, and MAP, NDCG@20, efficiency, and compression robustness are scored under identical inputs.

Load-bearing premise

The trade-off is assumed to be diagnosable from CSFCube’s small facet-rich computer-science set, RELISH’s larger biomedical overall ranking, and progressive deletion of fifty abstracts, while trainable deep unsupervised hashers stay outside the protocol.

What would settle it

On a larger multi-domain scientific corpus with controlled paraphrase and version-edit pairs, if a cheap lexical fingerprint matched or beat BGE-based codes on both near-duplicates and rewritten pairs under the same MAP and NDCG@20 protocol, the claimed consistent trade-off would fail.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Near-duplicate scientific cleaning can stay with cheap lexical or structural fingerprints when high surface overlap is expected.
  • Deduplication under paraphrase or summarization should prefer semantic quantization despite higher cost.
  • Hash methods must be reported with their compatible scorers, because some metric pairs collapse to the same ranking.
  • Facet-level method similarity remains harder than background or overall similarity for both hashing families.
  • Broader domains, stronger edits, and neighborhood-preserving quantizers on frozen encoders are the natural next stress tests the paper itself points to.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • A hybrid router that sends obvious near-duplicates to lexical hashes and hard rewrites to BGE codes could cut cost while keeping most ranking quality.
  • Rank-equivalence checks between scorers could become a standard reproducibility step in any hashing benchmark.
  • The protocol isolates plug-and-play generators; the full design space of trainable unsupervised hashers remains a separate question the results do not settle.
  • Method-level scientific similarity looking hardest on CSFCube suggests future labels should stress methodological overlap, not only topical relatedness.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 0 minor

Summary. H3D is a diagnostic benchmark for unsupervised text hashing aimed at fine-grained scientific document deduplication. It evaluates five non-learning fingerprint methods (MinHash, SimHash, Winnowing, FuzzyHash, FlyHash) and two frozen-BGE quantization pipelines (BGE-BIHash, BGE-LSHash) under a shared query–candidate ranking protocol on CSFCube (facet-level CS papers) and RELISH (larger biomedical overall-similarity ranking). The paper reports MAP and NDCG@20, runtime, method–scorer compatibility (including rank-equivalence collapses), a k-table ablation for BGE-LSHash, and a progressive LLM-deletion robustness probe on 50 CSFCube abstracts. The central claim is a trade-off: lexical/structural fingerprints are competitive for near-duplicate matching, while BGE-based representations better preserve similarity under content rewriting, at higher cost.

Significance. If the reported trade-off holds under the stated scope, H3D is a useful diagnostic resource for IR and scientific-document systems: it unifies heterogeneous hash outputs through type-specific adapters and scorers, documents when scorers become rank-equivalent, and makes accuracy–efficiency–robustness comparisons reproducible on public datasets. Strengths include the explicit exclusion of trainable deep unsupervised hashers (Table 2), the facet-level CSFCube analysis, the honest appendix diagnostics on zero runtime rows and rank-equivalence footprints, and the released code/dataset pointers. The contribution is primarily an evaluation protocol and empirical characterization rather than a new hashing architecture.

major comments (3)
  1. Abstract, §1, and §5.3/Fig. 8: the central claim that semantic-sensitive methods better preserve similarity under content rewriting is only weakly supported. RQ3 uses progressive LLM summarization/deletion (0–80%) on 50 CSFCube abstracts. That probe removes lexical evidence while partially preserving topic; it is not paraphrasing, translation, citation edits, or preprint–final variants that §2.2 and §6 themselves list as realistic scientific duplication modes. Without at least one meaning-preserving rewrite family of comparable length/structure, the rewriting half of the trade-off remains under-tested even if the ranking tables are sound.
  2. §3.1 and Table 2: excluding deep unsupervised semantic hashers (VDSH, NASH, BOTTLENECK, DHIM, MICPQ, etc.) is a defensible scope choice for isolating plug-and-play generators, but the abstract and conclusion still frame H3D as a general unsupervised text-hashing benchmark for fine-grained deduplication. The manuscript should either (i) restate the claim as limited to non-learning + frozen-embedding quantization, or (ii) add a minimal frozen/pretrained baseline comparison that does not require full task-specific training, so readers do not over-generalize the trade-off beyond the evaluated families.
  3. §5.1–5.2 and Tables 5–8: ranking quality is reported as point estimates without uncertainty (no query-level variance, bootstrap intervals, or multi-seed runs for stochastic components such as BGE-LSHash projections). Given CSFCube’s small query set (50) and the method–metric sensitivity shown in Figs. 4–5, some facet-level and best-configuration rankings may not be stable. At minimum, report query-level standard errors or multi-seed ranges for the headline MAP/NDCG comparisons that drive the trade-off claim.

Circularity Check

0 steps flagged

No circularity: empirical benchmark rankings on external public datasets with frozen encoders and handcrafted hashes; no fitted constants re-labeled as predictions.

full rationale

H3D is a diagnostic evaluation protocol, not a first-principles derivation. Load-bearing claims (lexical/structural hashes competitive for near-duplicates; BGE-based hashes more robust under compression, at higher cost) rest on MAP/NDCG@20 rankings of fixed-rule methods (MinHash, SimHash, Winnowing, FuzzyHash, FlyHash) and frozen BGE + post-hoc BIHash/LSHash quantizers on the public CSFCube and RELISH query–candidate pools, plus a controlled progressive-deletion probe. No parameters are fitted to the evaluation labels and then re-presented as predictions; BGE encoders are frozen and non-learning hashes use handcrafted rules. Rank-equivalence of scorers for binary/count signatures is diagnosed and reported rather than hidden. Self-citations, if any, are not load-bearing uniqueness theorems that force the trade-off. The paper is self-contained against external graded labels under a unified ranking interface, so the derivation chain does not reduce to its inputs by construction.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 2 invented entities

The central empirical trade-off rests on standard IR ranking metrics, two public scientific similarity collections, frozen BGE encoders, handcrafted hash rules, and several implementation defaults (bit length, n-gram size, seeds, LLM deletion ratios). No new physical entities are postulated; free parameters are engineering knobs that affect absolute scores but not the qualitative lexical-vs-semantic pattern.

free parameters (5)
  • BGE embedding dimension D and target code length L (main layout D=1024, L=32/64/128)
    Chosen as common defaults; reshape/pooling and LSH bit budget depend on them and change absolute MAP/NDCG.
  • LSH number of tables/subspaces k (grid 4–16, L≈128)
    Hyperparameter of BGE-LSHash; paper reports near-flat MAP but runtime still depends on k.
  • n-gram size 3, Winnowing window 5, FlyHash WTA top-8 / projection dim, random seeds 123/55
    Hand-set defaults that define the non-learning fingerprints and stochastic projections.
  • LLM progressive deletion ratios (0–80%) on 50 CSFCube documents
    Controls the robustness probe; different summarizer or sample would change the curves.
  • Binary relevance threshold g0=2 for MAP on graded labels
    Converts graded CSFCube/RELISH labels into binary relevance for AP; affects ranking scores.
axioms (5)
  • domain assumption Unsupervised non-learning hashes and frozen-embedding post-hoc quantization can be fairly compared without task-specific neural training on the benchmark corpora.
    Defines H3D scope in §3.1/Table 2; excludes VDSH/NASH/etc. so results do not speak to trained semantic hashers.
  • domain assumption CSFCube facet labels and RELISH graded overall labels are valid ground truth for fine-grained scientific similarity ranking.
    All MAP/NDCG claims rest on these annotations (§2).
  • ad hoc to paper Type-specific adapters (string/set/multiset/vector) plus chosen scorers implement a common ranking interface without unfairly favoring one family.
    §4.3; paper itself shows some scorers become rank-equivalent only for certain representations.
  • ad hoc to paper LLM-based progressive shortening preserves main topic while removing lexical evidence, thus probing semantic robustness.
    RQ3 setup in §5.3; not a general model of paraphrase or citation edit.
  • standard math Standard IR ranking metrics (MAP with binary threshold, NDCG@20) measure preservation of similarity under hashing.
    §4.2 definitions used throughout results.
invented entities (2)
  • H3D benchmark protocol (unified query–candidate pipeline + scorer-aligned hash comparison + compression probe) independent evidence
    purpose: Provide a diagnostic comparison framework for unsupervised scientific-document hashing families.
    The named contribution of the paper; independent evidence is the released datasets/artifacts and tabulated results, not an external physical measurement.
  • BGE-BIHash and BGE-LSHash quantizers independent evidence
    purpose: Turn frozen BGE dense vectors into compact binary fingerprints via pooling+sign or multi-subspace random hyperplanes.
    Lightweight post-hoc constructions built from known BGE + LSH/binarization ideas; not new particles or forces.

pith-pipeline@v1.1.0-grok45 · 31714 in / 3290 out tokens · 34366 ms · 2026-07-10T08:21:00.854005+00:00 · methodology

0 comments
read the original abstract

Document hashing provides compact representations for efficient similarity search and document deduplication, but existing studies rarely compare hashing pipelines under a unified protocol for fine-grained scientific documents. H3D is an unsupervised text hashing benchmark for fine-grained document deduplication. It evaluates representative unsupervised non-learning hashing approaches (MinHash, SimHash, Winnowing, FuzzyHash, FlyHash) together with semantic-sensitive methods built from frozen BGE embeddings and two quantization strategies (BGE-BIHash and BGE-LSHash). The non-learning methods generate hash fingerprints through manually designed mathematical rules without training or labeled similarity pairs, which distinguishes them from neural semantic hashing models. We benchmark all methods on CSFCube and RELISH, two datasets that provide complementary evaluation settings: facet-level analysis for scientific-document similarity and larger-scale split-level evaluation for biomedical similarity search. H3D jointly reports ranking quality (MAP, NDCG@20), efficiency, and robustness under controlled text compression. The results show a consistent trade-off: lexical and structural fingerprints are competitive for near-duplicate matching, while semantic-sensitive representations better preserve similarity under content rewriting, at higher computational cost. We further analyze when different similarity measures become rank-equivalent for specific hash representations, improving the interpretability and reproducibility of method comparisons.

Figures

Figures reproduced from arXiv: 2607.08382 by Bo Li, Hanwen Hao, Jiaxun Lyu, Jingzheng Li, Junnan Liu, Qianren Mao, Zhijun Chen.

Figure 1
Figure 1. Figure 1: Overview of best-performing hashing configurations on CSFCUBE and RELISH, generated from the ex [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Facet-level MAP comparison on CSFCUBE. Each row reports the best configuration for a hashing method, [PITH_FULL_IMAGE:figures/full_fig_p013_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Facet-level NDCG@20 comparison on CSFCUBE. Each row reports the best configuration for a hashing [PITH_FULL_IMAGE:figures/full_fig_p013_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Compatibility between hash methods and similarity measures on CSFCUBE. Values denote overall MAP, [PITH_FULL_IMAGE:figures/full_fig_p014_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Compatibility between hash methods and similarity measures on CSFCUBE. Values denote overall [PITH_FULL_IMAGE:figures/full_fig_p014_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: RELISH Test/Dev MAP comparison. Each row reports the best configuration for a hashing method according to Test MAP. 15 [PITH_FULL_IMAGE:figures/full_fig_p015_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: RELISH Test/Dev NDCG@20 comparison. The close Test/Dev values indicate that the ranking behavior is largely consistent across splits. 5.2 Semantic-Sensitive Hash Analysis (RQ2) We evaluate semantic-sensitive hashing methods using the BAAI General Embedding (BGE) family Xiao et al. (2023). Unlike unsupervised non-learning hashing methods, BGE-based hashing starts from contextual dense embed￾dings, which can… view at source ↗
Figure 8
Figure 8. Figure 8: Comparison of Semantic-Agnostic and Semantic-Sensitive Hashing Robustness on the CSFCUBE Dataset. The subfigures compare similarity changes under progressive text deletion for semantic-agnostic methods and BGE-based semantic-sensitive methods. Mapping: (a) Min￾Hash, (b) SimHash, (c) FlyHash, (d) FuzzyHash, (e) Winnowing, (f) BGE-BIHash, (g) BGE-base, (h) BGE-LSHash. 19 [PITH_FULL_IMAGE:figures/full_fig_p0… view at source ↗
Figure 9
Figure 9. Figure 9: Appendix diagnostics from released summary artifacts. (A) RELISH split-gap stability across configurations [PITH_FULL_IMAGE:figures/full_fig_p020_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Time–quality Pareto scatter on RELISH semantic-agnostic configurations from Table [PITH_FULL_IMAGE:figures/full_fig_p021_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: NDCG@K performance across different K values on RELISH (unfaceted) and CSFCUBE (facet=all) datasets. All results were obtained by averaging over three independent runs, each conducted on a 10% random sample of the development (dev) set. FuzzyHash+Jaccard BGE-LSHash(large) MinHash+Jaccard SimHash+Manhattan Winnowing+M-jaccard 0.0 0.1 0.2 0.3 0.4 0.5 0.6 Score (A) RELISH (unfaceted): P/R/F1@20 Precision@20 … view at source ↗
Figure 12
Figure 12. Figure 12: Extended metrics from aggregated evaluations for representative methods (Winnowing [PITH_FULL_IMAGE:figures/full_fig_p022_12.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

21 extracted references · 21 canonical work pages · 2 internal anchors

  1. [1]

    A.Z. Broder. On the resemblance and containment of documents. InProceedings. Compression and Complexity of SEQUENCES 1997 (Cat. No.97TB100171), pp. 21–29,

    work page 1997

  2. [2]

    Peter Brown and Yaoqi Zhou

    doi: 10.1109/SEQUEN.1997.666900. Peter Brown and Yaoqi Zhou. Large expert-curated database for benchmarking document similarity detection in biomedical literature search.Database, 2019:baz085,

    work page doi:10.1109/sequen.1997.666900 1997

  3. [3]

    URLhttps://doi.org/10.1145/3077136.3080816

    doi: 10.1145/3077136.3080816. URLhttps://doi.org/10.1145/3077136.3080816. Tobias Christiani and Rasmus Pagh. Set similarity search beyond minhash. InProceedings of the 49th Annual ACM SIGACT Symposium on Theory of Computing, STOC, pp. 1094–1107. ACM,

    work page doi:10.1145/3077136.3080816

  4. [4]

    URLhttps://doi.org/10.1145/3055399.3055443

    doi: 10.1145/3055399.3055443. URLhttps://doi.org/10.1145/3055399.3055443. Sanjoy Dasgupta, Charles F Stevens, and Saket Navlakha. A neural algorithm for a fundamental computing problem. Science, 358(6364):793–796,

    work page doi:10.1145/3055399.3055443

  5. [5]

    Doan and Chandan K

    27 Preprint Khoa D. Doan and Chandan K. Reddy. Efficient implicit unsupervised text hashing using adversarial autoencoder. In WWW ’20: The Web Conference 2020, Taipei, Taiwan, April 20-24, 2020, pp. 684–694. ACM/IW3C2,

    work page 2020

  6. [6]

    URLhttps://doi.org/10.1145/3366423.3380150

    doi: 10.1145/3366423.3380150. URLhttps://doi.org/10.1145/3366423.3380150. Wei Dong, Qinliang Su, Dinghan Shen, and Changyou Chen. Document hashing with mixture-prior generative mod- els. InProceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing, EMNLP-...

    work page doi:10.1145/3366423.3380150 2019

  7. [7]

    URLhttps://doi.org/10.18653/ v1/D19-1526

    doi: 10.18653/V1/D19-1526. URLhttps://doi.org/10.18653/ v1/D19-1526. Qianxin Huang, Siyao Peng, Xiaobo Shen, Yunhao Yuan, and Shirui Pan. Similarity preserving transformer cross- modal hashing for video-text retrieval. InProceedings of the 32nd ACM International Conference on Multime- dia, MM, pp. 5883–5891. ACM,

    work page doi:10.18653/v1/d19-1526

  8. [8]

    URLhttps://doi.org/10.1145/ 3664647.3681606

    doi: 10.1145/3664647.3681606. URLhttps://doi.org/10.1145/ 3664647.3681606. Piotr Indyk and Rajeev Motwani. Approximate nearest neighbors: towards removing the curse of dimensionality. In Proceedings of the thirtieth annual ACM symposium on Theory of computing, pp. 604–613,

    work page doi:10.1145/3664647.3681606

  9. [10]

    A Survey on Locality Sensitive Hashing Algorithms and their Applications

    URLhttps: //arxiv.org/abs/2102.08942. Qixia Jiang and Maosong Sun. Semi-supervised simhash for efficient document similarity search. InThe 49th Annual Meeting of the Association for Computational Linguistics: Human Language Technologies (ACL), pp. 93–101. The Association for Computer Linguistics,

    work page internal anchor Pith review Pith/arXiv arXiv

  10. [12]

    Jesse Kornblum

    URLhttps://arxiv.org/abs/2411.04257. Jesse Kornblum. Identifying almost identical files using context triggered piecewise hashing.Digital investigation, 3: 91–97,

    work page arXiv

  11. [13]

    Deduplicating training data makes language models better

    Katherine Lee, Daphne Ippolito, Andrew Nystrom, Chiyuan Zhang, Douglas Eck, Chris Callison-Burch, and Nicholas Carlini. Deduplicating training data makes language models better. In Smaranda Muresan, Preslav Nakov, and Aline Villavicencio (eds.),Proceedings of the 60th Annual Meeting of the Association for Computational Linguis- tics (Volume 1: Long Papers...

    work page 2022

  12. [14]

    URLhttps://doi.org/10.18653/ v1/2022.acl-long.577

    doi: 10.18653/V1/2022.ACL-LONG.577. URLhttps://doi.org/10.18653/ v1/2022.acl-long.577. Gurmeet Singh Manku, Arvind Jain, and Anish Das Sarma. Detecting near-duplicates for web crawling. InProceed- ings of the 16th international conference on World Wide Web, pp. 141–150,

    work page doi:10.18653/v1/2022.acl-long.577 2022

  13. [15]

    CSFCube -- A Test Collection of Computer Science Research Articles for Faceted Query by Example

    Sheshera Mysore, Tim O’Gorman, Andrew McCallum, and Hamed Zamani. Csfcube–a test collection of computer science research articles for faceted query by example.arXiv preprint arXiv:2103.12906,

    work page internal anchor Pith review Pith/arXiv arXiv

  14. [16]

    Veidenbaum, Alex Nicolau, and Tony Givargis

    Igor Nunes, Mike Heddes, Pere Vergés, Danny Abraham, Alexander V . Veidenbaum, Alex Nicolau, and Tony Givargis. Dothash: Estimating set similarity metrics for link prediction and document deduplication. InProceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, KDD 2023, Long Beach, CA, USA, August 6-10, 2023, pp. 1758–1769. ACM,

    work page 2023

  15. [17]

    URLhttps://doi.org/ 10.1145/3580305.3599314

    doi: 10.1145/3580305.3599314. URLhttps://doi.org/ 10.1145/3580305.3599314. Zijing Ou, Qinliang Su, Jianxing Yu, Ruihui Zhao, Yefeng Zheng, and Bang Liu. Refining BERT embeddings for document hashing via mutual information maximization. InFindings of the Association for Computational Lin- guistics: EMNLP, pp. 2360–2369. Association for Computational Linguistics,

    work page doi:10.1145/3580305.3599314

  16. [18]

    FINDINGS-EMNLP.203

    doi: 10.18653/V1/2021. FINDINGS-EMNLP.203. URLhttps://doi.org/10.18653/v1/2021.findings-emnlp.203. 28 Preprint Zexuan Qiu, Qinliang Su, Zijing Ou, Jianxing Yu, and Changyou Chen. Unsupervised hashing with contrastive information bottleneck. InProceedings of the Thirtieth International Joint Conference on Artificial Intelligence, IJCAI, pp. 959–965. ijcai.org,

    work page doi:10.18653/v1/2021 2021

  17. [19]

    URLhttps://doi.org/10.24963/ ijcai.2021/133

    doi: 10.24963/IJCAI.2021/133. URLhttps://doi.org/10.24963/ ijcai.2021/133. Zexuan Qiu, Qinliang Su, Jianxing Yu, and Shijing Si. Efficient document retrieval by end-to-end refining and quantizing BERT embedding with contrastive product quantization. InProceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, EMNLP, pp. 853–86...

    work page doi:10.24963/ijcai.2021/133 2021

  18. [20]

    URLhttps://doi.org/10.18653/v1/2022

    doi: 10.18653/V1/2022.EMNLP-MAIN.54. URLhttps://doi.org/10.18653/v1/2022. emnlp-main.54. BiChen Rao and Erkang Zhu. Searching web data using minhash LSH. InProceedings of the 2016 International Conference on Management of Data, SIGMOD, pp. 2257–2258. ACM,

    work page doi:10.18653/v1/2022.emnlp-main.54 2022

  19. [21]

    URLhttps://doi.org/10.1145/2882903.2914838

    doi: 10.1145/2882903.2914838. URLhttps://doi.org/10.1145/2882903.2914838. Saul Schleimer, Daniel S Wilkerson, and Alex Aiken. Winnowing: local algorithms for document fingerprinting. In Proceedings of the 2003 ACM SIGMOD international conference on Management of data, pp. 76–85, 2003a. Saul Schleimer, Daniel Shawcross Wilkerson, and Alex Aiken. Winnowing:...

    work page doi:10.1145/2882903.2914838 2003

  20. [22]

    URLhttps://aclanthology.org/ P18-1190/

    doi: 10.18653/V1/P18-1190. URLhttps://aclanthology.org/ P18-1190/. Shitao Xiao, Zheng Liu, Peitian Zhang, and Niklas Muennighoff. C-pack: Packaged resources to advance general chinese embedding,

    work page doi:10.18653/v1/p18-1190

  21. [23]

    URL https://doi.org/10.18653/v1/2020.acl-main.71

    doi: 10.18653/V1/2020.ACL-MAIN.71. URL https://doi.org/10.18653/v1/2020.acl-main.71. 29

    work page doi:10.18653/v1/2020.acl-main.71 2020