pith. machine review for the scientific record. sign in

arxiv: 2605.01073 · v1 · submitted 2026-05-01 · 💻 cs.CL

Recognition: unknown

Controlled Paraphrase Geometry in Sentence Embedding Space: Local Manifold Modeling and Latent Probing

Leonid Bedratyuk
Authors on Pith no claims yet

Pith reviewed 2026-05-09 18:49 UTC · model grok-4.3

classification 💻 cs.CL
keywords sentence embeddingslocal manifold modelingparaphrase geometrylatent probingcontrolled datasetembedding space geometrysurface fittingnonlinear models
0
0 comments X

The pith

Nonlinear low-degree surfaces model paraphrase-induced embedding clouds more accurately than affine planes, with surface-based probing preserving geometric consistency.

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

The paper examines how controlled semantic variations from near-paraphrases form local clouds in sentence embedding space and tests whether these clouds can be explicitly captured by simple fitted surfaces. It introduces affine, quadratic, and cubic models plus a probing method that generates new latent points on the fitted surface in a reduced PCA subspace. Experiments on template-controlled sentence sets demonstrate that quadratic and cubic fits describe the data shape better than linear ones while keeping neighborhood relations and second-order descriptors stable. The work also releases a 300K-sentence controlled dataset with slot annotations to support such geometry studies. Geometric fidelity of the generated points, however, does not translate into automatic gains on downstream classification tasks.

Core claim

Controlled local classes of semantically close sentences induce embedding clouds whose local structure is better described by quadratic and cubic fitted carriers than by affine ones; a surface-based latent probing procedure that constructs synthetic points relative to the fitted carrier yields high consistency on surface adherence, Hessian shape descriptors, coefficient stability, and empirical distribution agreement.

What carries the argument

Local geometric modeling scheme that fits affine, quadratic, and cubic surfaces to embedding clouds and performs surface-based latent probing by constructing synthetic points in a reduced local PCA space anchored to the fitted carrier.

If this is right

  • Nonlinear local models capture the shape of paraphrase embedding clouds with greater accuracy than linear ones.
  • Surface-based generation of latent points maintains high fidelity on surface consistency, Hessian-based shape descriptors, and fitted coefficients.
  • Geometric validity of synthetic points does not automatically produce better performance on downstream classification tasks.
  • Explicit local manifold modeling is a viable offline tool for analyzing sentence embedding geometry.
  • Controlled template-based datasets enable reproducible study of local embedding structure.

Where Pith is reading between the lines

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

  • The same fitting and probing pipeline could be applied to other controlled variation types such as syntactic transformations or lexical substitutions to test generality.
  • If natural paraphrases exhibit similar polynomial structure, the method might offer a way to regularize embedding spaces for consistency without task-specific supervision.
  • The observed gap between geometric fidelity and discriminative utility suggests that manifold-aware augmentation may require additional alignment steps to benefit classifiers.
  • Hessian stability metrics could serve as a diagnostic for whether an embedding model has learned locally smooth semantic neighborhoods.

Load-bearing premise

Template-controlled paraphrases generate embedding clouds whose local structure is faithfully represented by low-degree polynomial surfaces without introducing systematic artifacts absent from natural semantic variation.

What would settle it

Finding that quadratic and cubic models show no statistically significant improvement over affine models in fit quality or consistency metrics when the same procedure is applied to naturally occurring paraphrases collected without templates.

Figures

Figures reproduced from arXiv: 2605.01073 by Leonid Bedratyuk.

Figure 1
Figure 1. Figure 1: Visual comparison of representative local embedding clouds and fitted quadratic surfaces. Top row: pairwise PCA(3) projections of the C-C5 cloud, illustrating its clearly nonlinear structure. Bottom row: PCA(3) clouds for A-C5, B-C5, and C-C5 together with fitted quadratic surfaces shown for visualization purposes. and the fitted quadric reproduces their general shape for different template families. This … view at source ↗
Figure 2
Figure 2. Figure 2: Validation RMSE across representative local classes and geometric models. Lower values indicate better approximation quality on held-out data. Affine models are consistently weaker, while quadratic and cubic models capture the local geometry substantially better. the fitted local carrier. Others characterize empirical-density fidelity, that is, the closeness of synthetic points to the empirical density of … view at source ↗
Figure 3
Figure 3. Figure 3: Comparison of synthetic latent points generated by linear interpola￾tion, local random perturbation, and the proposed surface-based method for one representative local class in the PCA(3) space. 5.3. Geometric validity and downstream discriminative utility. The third experiment was aimed at examining the relation between two different properties of synthetic latent points: their geometric validity with res… view at source ↗
read the original abstract

The paper studies the local geometry of embedding clouds induced by \emph{controlled local classes of semantically close sentences}. The central question is how controlled paraphrase-like semantic variation is organized in sentence embedding space and whether this local structure can be explicitly modeled by low-degree fitted carriers. We introduce a local geometric modeling scheme based on affine, quadratic, and cubic fitted models. We also use a surface-based latent probing procedure that constructs synthetic latent points in a reduced local PCA space with respect to the fitted carrier. The procedure is intended as an offline method for representation-space analysis, local manifold modeling, and geometry-aware latent probing. Generated latent points are evaluated using criteria that measure consistency with the fitted surface, preservation of neighborhood structure, agreement with the empirical distribution, stability of Hessian-based second-order shape descriptors, and stability of fitted-model coefficients. Experiments on controlled sets of semantically close sentences show that nonlinear local models describe embedding clouds more accurately than affine models. Surface-based generation provides strong fitted-geometry fidelity, including surface consistency, Hessian-based shape consistency, and coefficient consistency. Downstream experiments show that geometric validity of synthetic latent points does not automatically translate into improved classification performance. The results support explicit local geometric modeling of sentence embedding space and highlight the need to distinguish geometric validity from discriminative utility. As a resource contribution, we introduce \textbf{CoPaGE-300K}, a controlled template-based dataset of semantically close sentence variants with slot-level annotations and precomputed sentence embeddings.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The paper claims that local geometry in sentence embedding space for controlled sets of semantically close sentences (via the new CoPaGE-300K template-based dataset) is better captured by fitted nonlinear (quadratic and cubic) local models than by affine models. It introduces a surface-based latent probing procedure that generates synthetic points in a reduced local PCA space relative to the fitted carrier and evaluates them on criteria including surface consistency, Hessian-based shape stability, coefficient stability, neighborhood preservation, and empirical distribution agreement. Experiments reportedly show strong fidelity for surface-based generation under nonlinear models, but geometric validity of the synthetic points does not translate to improved downstream classification performance.

Significance. If the central results hold without template-induced artifacts, the work supplies an explicit, offline framework for local manifold modeling and geometry-aware probing in embedding spaces, together with a large controlled resource (CoPaGE-300K) that enables slot-level analysis of paraphrase variation. It usefully separates geometric fidelity from discriminative utility, which could inform future representation analysis and controlled generation methods.

major comments (2)
  1. [§5 (Experiments) and Abstract] §5 (Experiments) and Abstract: the central claim that nonlinear local models describe embedding clouds more accurately than affine models is demonstrated exclusively on template-controlled paraphrases from CoPaGE-300K. Slot-filling generation can introduce low-degree algebraic dependencies that quadratic and cubic carriers fit by construction while affine models cannot; no ablation replacing the templates with organic paraphrase variation (back-translation, human rewrites, or existing paraphrase corpora) is reported to test whether the observed Hessian consistency, coefficient stability, and surface fidelity survive.
  2. [Abstract and §5] Abstract and §5: the statement that nonlinear models outperform affine ones and that generated points satisfy the geometric criteria is given without any quantitative metrics, error bars, statistical tests, or full experimental protocol details (e.g., exact model degrees, PCA dimensionality, fitting procedure, or number of local neighborhoods). This prevents assessment of effect size and reproducibility.
minor comments (2)
  1. [§3] Notation for the local affine/quadratic/cubic carriers and the surface-based generation procedure would benefit from explicit equations (e.g., the precise form of the fitted polynomial and the projection onto the reduced PCA basis) in §3.
  2. The paper should add a limitations paragraph discussing the extent to which template-induced structure may affect generalization to natural language variation.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their thorough review and valuable feedback on our manuscript. We address the major comments point by point below and outline the revisions we plan to make.

read point-by-point responses

  1. Referee: §5 (Experiments) and Abstract: the central claim that nonlinear local models describe embedding clouds more accurately than affine models is demonstrated exclusively on template-controlled paraphrases from CoPaGE-300K. Slot-filling generation can introduce low-degree algebraic dependencies that quadratic and cubic carriers fit by construction while affine models cannot; no ablation replacing the templates with organic paraphrase variation (back-translation, human rewrites, or existing paraphrase corpora) is reported to test whether the observed Hessian consistency, coefficient stability, and surface fidelity survive.

    Authors: The use of template-based paraphrases in CoPaGE-300K is central to our study because it provides precise control over the semantic variation through slot-filling, enabling a clean analysis of local manifold structure in embedding space without extraneous variations from organic rewrites. This design isolates the geometric effects of controlled paraphrase classes, which is the explicit focus of the work. While organic paraphrase data could offer complementary validation, it would introduce confounding syntactic and lexical factors that obscure the local manifold analysis we target. We will add a discussion section in the revision explaining the rationale for templates and acknowledging the limitation on generalizability to natural data. revision: partial

  2. Referee: Abstract and §5: the statement that nonlinear models outperform affine ones and that generated points satisfy the geometric criteria is given without any quantitative metrics, error bars, statistical tests, or full experimental protocol details (e.g., exact model degrees, PCA dimensionality, fitting procedure, or number of local neighborhoods). This prevents assessment of effect size and reproducibility.

    Authors: The full experimental protocol details, including model degrees (affine/quadratic/cubic), local PCA dimensionality, fitting procedures, and neighborhood counts, are described in §5. However, we agree that the abstract and summary lack explicit quantitative metrics, error bars, and statistical tests. In the revision we will update the abstract with key quantitative results (e.g., mean surface-consistency and stability scores with standard deviations), add error bars to figures, and report statistical comparisons to quantify effect sizes and support reproducibility. revision: yes

Circularity Check

0 steps flagged

No significant circularity detected

full rationale

The paper fits affine/quadratic/cubic carriers to embedding clouds from the new CoPaGE-300K template dataset, then evaluates synthetic points generated from those carriers. While surface-consistency checks are partly tautological by construction, the central claim (nonlinear models describe the clouds more accurately) rests on comparative fit quality plus independent external criteria such as neighborhood preservation and agreement with the empirical distribution. No self-citations, uniqueness theorems, or ansatzes imported from prior author work appear; the derivation chain does not reduce to its inputs by definition.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The central claim depends on the assumption that local embedding clouds admit low-degree polynomial approximations and that the listed geometric criteria are meaningful proxies for manifold fidelity. Free parameters are the coefficients of the fitted affine/quadratic/cubic models. No new physical entities are postulated.

free parameters (1)
  • coefficients of local affine/quadratic/cubic models
    Fitted directly to the reduced PCA coordinates of each paraphrase cloud; central to all accuracy and consistency claims.
axioms (1)
  • domain assumption Local clouds of controlled paraphrases form low-dimensional structures that low-degree polynomials can approximate without significant residual structure.
    Invoked when choosing affine, quadratic, and cubic carriers and when interpreting Hessian stability as a shape descriptor.

pith-pipeline@v0.9.0 · 5562 in / 1383 out tokens · 53530 ms · 2026-05-09T18:49:52.903643+00:00 · methodology

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

38 extracted references · 28 canonical work pages · 1 internal anchor

  1. [1]

    Computational Linguistics, 49(2):465–523, 2023

    Apidianaki, M.From Word Types to Tokens and Back: A Survey of Approaches to Word Meaning Rep- resentation and Interpretation. Computational Linguistics, 49(2):465–523, 2023. DOI:https://doi.org/ 10.1162/coli_a_00474

    work page doi:10.1162/coli_a_00474 2023

  2. [2]

    Computational Linguistics, 47(3):663–698, 2021

    Chersoni, E., Santus, E., Huang, C.-R., and Lenci, A.Decoding Word Embeddings with Brain-Based Se- mantic Features. Computational Linguistics, 47(3):663–698, 2021. DOI:https://doi.org/10.1162/coli_ a_00412

    work page doi:10.1162/coli_ 2021

  3. [3]

    Transactions of the Association for Computational Linguistics, 10:573–588, 2022

    Sun, X., Meng, Y., Ao, X., Wu, F., Zhang, T., Li, J., and Fan, C.Sentence Similarity Based on Contexts. Transactions of the Association for Computational Linguistics, 10:573–588, 2022. DOI:https://doi.org/ 10.1162/tacl_a_00477

    work page doi:10.1162/tacl_a_00477 2022

  4. [4]

    Transactions of the Association for Computational Linguistics, 11:997–1013,

    Wang, Y., Tao, S., Xie, N., Yang, H., Baldwin, T., and Verspoor, K.Collective Human Opinions in Semantic Textual Similarity. Transactions of the Association for Computational Linguistics, 11:997–1013,

  5. [5]

    DOI:https://doi.org/10.1162/tacl_a_00584

    work page doi:10.1162/tacl_a_00584

  6. [6]

    Computational Linguistics, 51(1):139–190, 2025

    Fodor, J., De Deyne, S., and Suzuki, S.Compositionality and Sentence Meaning: Comparing Seman- tic Parsing and Transformers on a Challenging Sentence Similarity Dataset. Computational Linguistics, 51(1):139–190, 2025. DOI:https://doi.org/10.1162/coli_a_00536

    work page doi:10.1162/coli_a_00536 2025

  7. [7]

    A.Information Theory–based Compositional Distributional Semantics

    Amig´ o, E., Ariza-Casabona, A., Fresno, V., and Mart´ ı, M. A.Information Theory–based Compositional Distributional Semantics. Computational Linguistics, 48(4):907–948, 2022. DOI:https://doi.org/10. 1162/coli_a_00454

    2022

  8. [8]

    Vuli´ c, I., Baker, S., Ponti, E. M., Petti, U., Leviant, I., Wing, K., Majewska, O., Bar, E., Malone, M., Poibeau, T., Reichart, R., and Korhonen, A.Multi-SimLex: A Large-Scale Evaluation of Multilingual and Crosslingual Lexical Semantic Similarity. Computational Linguistics, 46(4):847–897, 2020. DOI:https: //doi.org/10.1162/coli_a_00391. 36 LEONID BEDRATYUK

    work page doi:10.1162/coli_a_00391 2020

  9. [9]

    and Dagan, I.Still a pain in the neck: Evaluating Text Representations on Lexical Composition

    Shwartz, V. and Dagan, I.Still a pain in the neck: Evaluating Text Representations on Lexical Composition. Transactions of the Association for Computational Linguistics, 7:403—419., 2019. DOI:https://doi.org/ 10.1162/tacl_a_00277

    work page doi:10.1162/tacl_a_00277 2019

  10. [10]

    Semantics of Multiword Expressions in Transformer-Based Models: A Survey

    Mileti´ c, F. and Schulte im Walde, S.Semantics of Multiword Expressions in Transformer-Based Language Models. Transactions of the Association for Computational Linguistics, 12:593–612, 2024. DOI:https: //doi.org/10.1162/tacl_a_00657

    work page doi:10.1162/tacl_a_00657 2024

  11. [11]

    Transactions of the Association for Computational Linguistics, 12:299–320,

    Gar´ ıSoler, A., Labeau, M., andClavel, C.The Impact of Word Splitting on the Semantic Content of Contex- tualized Word Representations. Transactions of the Association for Computational Linguistics, 12:299–320,

  12. [12]

    DOI:https://doi.org/10.1162/tacl_a_00647

    work page doi:10.1162/tacl_a_00647

  13. [13]

    Holmes: A Benchmark to Assess the Linguistic Competence of Language Models

    Waldis, A., Perlitz, Y., Choshen, L., Hou, Y., and Gurevych, I.Holmes: A Benchmark to Assess the Linguistic Competence of Language Models. Transactions of the Association for Computational Linguistics, 12:1616–1647, 2024. DOI:https://doi.org/10.1162/tacl_a_00718

    work page doi:10.1162/tacl_a_00718 2024

  14. [14]

    A., and Reuter, C

    Bayer, M., Kaufhold, M. A., and Reuter, C. (2022). A survey on data augmentation for text classification. ACM Computing Surveys, 55(7), 1-39. DOI:https://doi.org/10.1145/3544558

    work page doi:10.1145/3544558 2022

  15. [15]

    Contextual

    S. Kobayashi, “Contextual Augmentation: Data Augmentation by Words with Paradigmatic Relations,” in Proceedings of NAACL-HLT, 2018, pp. 452–457. DOI:https://doi.org/10.18653/v1/N18-2072

    work page doi:10.18653/v1/n18-2072 2018

  16. [16]

    EDA : Easy Data Augmentation Techniques for Boosting Performance on Text Classification Tasks

    J. Wei and K. Zou, “EDA: Easy Data Augmentation Techniques for Boosting Performance on Text Classi- fication Tasks,” inProceedings of the 2019 Conference on Empirical Methods in Natural Language Process- ing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), Hong Kong, China, 2019, pp. 6382–6388. DOI:https://doi.org/...

    work page doi:10.18653/v1/d19-1670 2019

  17. [17]

    How contextual are contextualized word representations? comparing the geometry of BERT, ELMo, and GPT-2 embeddings

    K. Ethayarajh, “How Contextual are Contextualized Word Representations? Comparing the Geometry of BERT, ELMo, and GPT-2 Embeddings,” inProceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP),HongKong, China, 2019, pp.55–65.DOI:https://do...

    work page doi:10.18653/v1/d19-1006 2019

  18. [18]

    On the Sentence Embeddings from Pre-trained Language Models

    B. Li, H. Zhou, J. He, M. Wang, Y. Yang, and L. Li, “On the Sentence Embeddings from Pre-trained Lan- guageModels,” inProceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), Online, 2020, pp. 9119–9130. DOI:https://doi.org/10.18653/v1/2020.emnlp-main.733

    work page doi:10.18653/v1/2020.emnlp-main.733 2020

  19. [19]

    Zanatta, A

    Y. Chu, H. Cao, Y. Diao, and H. Lin, “Refined SBERT: Representing sentence BERT in manifold space,”Neurocomputing, vol. 555, Article 126453, 2023. DOI:https://doi.org/10.1016/j.neucom.2023. 126453

    work page doi:10.1016/j.neucom.2023 2023

  20. [20]

    Semantic Geometry of Sentence Embeddings,

    M. Tehenan, “Semantic Geometry of Sentence Embeddings,” inFindings of the Association for Compu- tational Linguistics: EMNLP 2025, Suzhou, China, 2025, pp. 11993–12004. DOI:https://doi.org/10. 18653/v1/2025.findings-emnlp.641

    2025

  21. [21]

    UMAP: Uniform Manifold Approximation and Projection for Dimension Reduction

    L. McInnes, J. Healy, and J. Melville, “UMAP: Uniform Manifold Approximation and Projection for Di- mension Reduction,”arXiv preprint arXiv:1802.03426, 2018. URL:https://arxiv.org/abs/1802.03426

    work page internal anchor Pith review arXiv 2018

  22. [22]

    Sainburg, L

    T. Sainburg, L. McInnes, and T. Q. Gentner, “Parametric UMAP Embeddings for Representation and Semisupervised Learning,”Neural Computation, vol. 33, no. 11, pp. 2881–2907, 2021. DOI:https://doi. org/10.1162/neco_a_01434

    work page doi:10.1162/neco_a_01434 2021

  23. [23]

    Data augmentation: A comprehensive survey of modern approaches,

    A. Mumuni and F. Mumuni, “Data augmentation: A comprehensive survey of modern approaches,”Array, vol. 16, Article 100258, 2022. DOI:https://doi.org/10.1016/j.array.2022.100258

    work page doi:10.1016/j.array.2022.100258 2022

  24. [24]

    How Effective is Task-Agnostic Data Augmentation for Pretrained Transformers?

    S. Longpre, Y. Wang, and C. DuBois, “How Effective is Task-Agnostic Data Augmentation for Pretrained Transformers?” inFindings of the Association for Computational Linguistics: EMNLP 2020, Online, 2020, pp. 4401–4411. DOI:https://doi.org/10.18653/v1/2020.findings-emnlp.394

    work page doi:10.18653/v1/2020.findings-emnlp.394 2020

  25. [26]

    URL:https://arxiv.org/abs/1702.05538

    work page arXiv

  26. [27]

    Data Augmentation as Feature Manipulation,

    R. Shen, S. Bubeck, and S. Gunasekar, “Data Augmentation as Feature Manipulation,” inProceedings of the 39th International Conference on Machine Learning (ICML),PMLR 162:19773–19808 2022. URL: https://proceedings.mlr.press/v162/shen22a.html

    2022

  27. [28]

    mixup: Beyond Empirical Risk Minimization,

    H. Zhang, M. Cisse, Y. N. Dauphin, and D. Lopez-Paz, “mixup: Beyond Empirical Risk Minimization,” inInternational Conference on Learning Representations (ICLR), 2018. URL:https://openreview.net/ forum?id=r1Ddp1-Rb

    2018

  28. [29]

    Manifold Mixup: Better Representations by Interpolating Hidden States,

    V. Verma, A. Lamb, C. Beckham, A. Najafi, I. Mitliagkas, A. Courville, D. Lopez-Paz, and Y. Bengio, “Manifold Mixup: Better Representations by Interpolating Hidden States,” inProceedings of the 36th International Conference on Machine Learning, 2019, pp. 6438–6447. URL:https://proceedings.mlr. press/v97/verma19a.html. MANIFOLD MODELING 37

    2019

  29. [30]

    MixUp as Locally Linear Out-of-Manifold Regularization,

    H. Guo, Y. Mao, and R. Zhang, “MixUp as Locally Linear Out-of-Manifold Regularization,” inProceedings of the AAAI Conference on Artificial Intelligence, vol. 33, no. 1, 2019, pp. 3714–3722. DOI:https://doi. org/10.1609/aaai.v33i01.33013714

    work page doi:10.1609/aaai.v33i01.33013714 2019

  30. [31]

    and Carlsson, Marcel , month = apr, year =

    Y. El-Laham, E. Fons, D. Daudert, and S. Vyetrenko, “Augment on Manifold: Mixup Regularization with UMAP,” inProceedings of the IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2024, pp. 7040–7044. DOI:https://doi.org/10.1109/ICASSP48485.2024.10446585

    work page doi:10.1109/icassp48485.2024.10446585 2024

  31. [32]

    On-Manifold Adversarial Data Augmentation Improves Uncertainty Calibration,

    K. Patel, W. Beluch, D. Zhang, M. Pfeiffer, and B. Yang, “On-Manifold Adversarial Data Augmentation Improves Uncertainty Calibration,”arXiv preprint arXiv:1912.07458, 2019. URL:https://arxiv.org/ abs/1912.07458

    work page arXiv 1912

  32. [33]

    Data Augmen- tation with Manifold Exploring Geometric Transformations for Increased Performance and Robustness,

    M. Paschali, W. Simson, A. G. Roy, M. F. Naeem, R. G¨ obl, C. Wachinger, and N. Navab, “Data Augmen- tation with Manifold Exploring Geometric Transformations for Increased Performance and Robustness,” in Information Processing in Medical Imaging, Lecture Notes in Computer Science, vol. 11492, 2019, pp. 517–

    2019

  33. [34]

    DOI:https://doi.org/10.1007/978-3-030-20351-1_40

    work page doi:10.1007/978-3-030-20351-1_40

  34. [35]

    2023 , url =

    Y.-B. Moon, J. Kim, H. Kim, K. Son, and T.-H. Oh, “TextManiA: Enriching Visual Feature by Text-driven Manifold Augmentation,” inProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023, pp. 2526–2537. DOI:https://doi.org/10.1109/ICCV51070.2023.00239

    work page doi:10.1109/iccv51070.2023.00239 2023

  35. [36]

    What Is a Paraphrase?

    R. Bhagat and E. Hovy, “What Is a Paraphrase?”Computational Linguistics, vol. 39, no. 3, pp. 463–472,

  36. [37]

    DOI:https://doi.org/10.1162/COLI_a_00166

    work page doi:10.1162/coli_a_00166

  37. [38]

    small classes

    Bedratyuk, L.CoPaGE-300K: Controlled Paraphrase Geometry Embeddings. Zenodo, 2026. Version 1.0.0. URL:https://doi.org/10.5281/zenodo.19827690. AppendixA.Lexical v ariants used to construct controlled classes This appendix provides the complete lists of lexical variants used to construct controlled sets of semantically close sentences. Including these list...

    work page doi:10.5281/zenodo.19827690 2026

  38. [39]

    Ther= 20truncation preserves0.6978,0.6791, and0.7005of the total PCA variance forA-C5,B-C5, andC-C5, respectively, whereas the90%-variance criterion selectsr= 38, r= 39, andr= 38

    This setting is intentionally lighter than the full90%-variance reduced spaces reported in Table 6. Ther= 20truncation preserves0.6978,0.6791, and0.7005of the total PCA variance forA-C5,B-C5, andC-C5, respectively, whereas the90%-variance criterion selectsr= 38, r= 39, andr= 38. MANIFOLD MODELING 41 Table 6.Adaptive PCA dimensionality forall-MiniLM-L6-v2s...