Pith. sign in

REVIEW 5 major objections 4 minor 32 references

Named Entity Swapping for Metadata Anonymization in a Text Corpus

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

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

Pith's one-line read This paper claims that swapping named entities between semantically similar chunks of different documents breaks the quasi-identifier combinations that let LLMs attribute texts to their true metadata.

desk verdict A genuinely novel text-swapping scheme with a careful but narrow LLM benchmark, whose central claim is undercut by the 4% coverage and an overclaiming abstract. read the letter →

arxiv 2505.21128 v1 pith:3UGM53II submitted 2025-05-27 stat.AP

Jan Greve , Lukas Sablica This is my paper
classification stat.AP MSC 62D0562H30
keywords namedentityswappingmetadataanonymizationstatisticaldisclosurecontroldataquasi-identifierstextembeddingssphericalclusteringearningscalltranscripts
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

The paper proposes an anonymization procedure, named entity swapping, that stops large language models from attributing a text to its metadata—in the application, the company behind an earnings call transcript. Each text chunk is treated as a cell in a contingency table of named-entity categories, and entities of the same category are randomly interchanged between semantically similar chunks taken from different documents, severing the combinations that act as quasi-identifiers. On 35 information-sector earnings calls from Q1 2018, company-name prediction accuracy on swapped chunks fell from 0.888 to 0.564 for GPT-4o and from 0.934 to 0.602 for Gemini 2.5 Flash, while most words in the transcripts stayed intact. A decision-theoretic layer over swap rate, entity-category choices, and clustering constraints lets a user trade disclosure risk against data utility.

What carries the argument

The object that carries the method is the chunk-level contingency table $z$: each chunk becomes a cell indexed by the entities it contains in categories such as Organization, Person, Product, Location, Event, and Date, and a cell with exactly one chunk is a risky sample unique. Pairing uses embeddings $x_i \in S^{d-1}$ clustered by a spherical mixture model; within a cluster, pairs from different documents receive uniform swap weights. Interchanging entities of selected categories scrambles quasi-identifier combinations while preserving most words. Risk is quantified through the Ewens-Pitman estimate of population uniques, $\hat{S}_1 \approx \frac{\Gamma(\theta+1)}{\Gamma(\theta+\alpha)} N^\alpha$, and utility through the log-likelihood ratio of post-swap embeddings under the same mixture, with a linear trade-off selecting the optimal release.

What would settle it

Mask all named entities remaining in the post-swap chunks and re-run the same temperature-zero LLM attribution prompt on the 35-company corpus. If accuracy stays well above the roughly 1-in-35 chance level, disclosure is flowing through non-entity signals and the central claim fails; if accuracy falls to chance, the named-entity channel is confirmed as the operative one.

Watch

Extended reading notes

Core claim

The central claim is that breaking named-entity combinations—not suppressing whole documents or summarizing them—is enough to anonymize a text corpus against LLM metadata discovery. The empirical demonstration shows that after swapping, both tested LLMs lose a substantial fraction of their ability to name the true company, and McNemar's test rejects the hypothesis that pre- and post-swap classifications are equally good. The swap preserves most of the text and the marginal distributions of unswapped entity categories, so the corpus retains much of its statistical content; the method is therefore presented as a less information-destructive alternative to suppression-based anonymization. In the optimal release identified by the risk-utility frontier, Organization and Product entities are swapped at the maximum feasible swap rate.

Load-bearing premise

The load-bearing premise is that a text's metadata can be discovered only through combinations of named entities within its chunks; if style, financial figures, or words the entity recognizer misses can reveal the source, swapping entities alone will not anonymize the corpus.

Editorial extensions

If this is right

  • If LLM look-ahead bias in earnings-call analysis comes from the model knowing which company spoke, then anonymized chunks can be fed into sentiment, risk, or return-prediction tasks with that channel blocked.
  • Because the method preserves most words, it can replace summarization in applications that need the original statistical texture of the corpus.
  • The same workflow transfers to any corpus of structurally similar documents—for instance, medical transcripts across hospitals or patent filings across a sector—whose named-entity combinations disclose the source.
  • The swap rate, the choice of entity categories, and cluster constraints give data holders an explicit privacy-utility dial, so the same machinery can serve both low- and high-utility regimes.
  • The evaluation design—temperature-zero LLM attribution before and after on the exact swapped chunks—can be reused as a standard test for other text anonymization methods.

Reading between the lines

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

  • An untested stress test: after swapping, mask every remaining named entity and re-run the attribution task; any accuracy well above the roughly 1-in-35 chance level would show that style, numbers, or untagged terms also disclose the source and would bound the method's ceiling.
  • The paper's failure examples suggest a formal success criterion the authors leave implicit: a chunk is anonymized when an LLM's attribution mass is split across at least two companies rather than merely shifted to the wrong one.
  • A direct extension would be to pair chunks by entity-set similarity in addition to embedding similarity, so chunks with matching entity counts are paired; this targets the observed failure where unequal entity counts leave a tell-tale entity behind.
  • Because the utility measure is computed with the same embedding model used for clustering, a downstream task that relies on different text statistics might suffer more distortion than the utility measure reports; evaluating swapped corpora on downstream tasks would test this.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

5 major / 4 minor

Summary. The manuscript proposes 'named entity swapping' as a text anonymization method: chunks are clustered by embedding similarity, pairs from different documents in the same cluster are selected, and named entities (e.g., Organization and Product) are interchanged. The method is placed in a decision-theoretic framework with a risk measure based on population uniques and a utility measure based on the fitted spherical mixture. The application uses 35 Q1 2018 information-sector earnings-call transcripts (1819 chunks) and reports that GPT-4o and Gemini 2.5 Flash company-name prediction accuracy drops from about 0.89/0.93 to 0.56/0.60 on the ~4% of chunks that received swaps, with McNemar tests significant across 30 Monte Carlo draws.

Significance. If the central claims held, this would be a useful addition to statistical disclosure control for text: it adapts a classical method (data swapping) to unstructured text using embeddings, and it provides an external LLM-based evaluation rather than relying only on internal measures. The empirical setup is a strength: 30 Monte Carlo draws, reported standard deviations, and paired McNemar tests. However, the current evidence supports only a decrease in identifiability for swapped chunks; the corpus-level 'prevention' claim and the calibration claims require further support. The work is therefore promising but not yet ready as stated.

major comments (5)
  1. [Section 3.2.2 / Abstract / Section 4] The only LLM evaluation covers the rn ~ 72 chunks that received swaps (about 4% of n = 1819), and Table 2 reports post-swap accuracies of 0.564 (GPT-4o) and 0.602 (Gemini), which are still far above chance for 35 companies. The abstract's 'prevents the disclosure of company names' and Section 4's 'no longer able to guess the correct company names' therefore overstate the evidence; the data support a significant decrease on swapped chunks only. Please either test unswapped chunks as well (e.g., a random sample of unchanged chunks with the same prompt) and report corpus-level accuracy, or explicitly scope all claims to swapped chunks and to risk reduction rather than prevention.
  2. [Section 2.5, Eq. (2.17)] Equation (2.17) does not define a proper likelihood ratio: the right-hand side evaluates the pre-swap embedding x'_i under the component indexed by the post-swap cluster assignment, while the left-hand side conditions on l, the membership vector for all observations. If a pre-swap embedding moves from cluster A to cluster B, the equation compares log f(˜x'_i|φ_B) with log f(x'_i|φ_B) instead of with log f(x'_i|φ_A). This can bias the utility measure and makes Eq. (2.19) unreliable as a measure of utility loss. Please correct the conditioning, for example by using the full mixture log-likelihood for both pre- and post-swap data, or by explicitly defining a joint cluster assignment.
  3. [Sections 2.2 and 2.5, Eq. (2.19)] DU is defined with the same fitted spherical mixture used in Section 2.2 to generate the swapping weights. Consequently, the utility measure partly certifies its own pairing choices: a pair is likely to receive high utility exactly when the post-swap embedding stays near a cluster center that the same fitted model selected. The external LLM check mitigates this for the effectiveness claim, but not for the calibration/utility claim. Please validate DU against an independent utility measure, such as downstream task performance or an embedding model not used for pairing.
  4. [Section 3.2.1, Figure 3] The identity of the best swapping category combination is not robust to the superpopulation size N. Product+Location lies on the efficient frontier for N = 1e10 but is the worst combination for N = 1e20 and N = 1e30, with the ranking driven by N^{α-1} in Eq. (2.15). The paper selects N via 'conservative' bounds rather than estimation, yet this choice changes the recommended release. The decision-theoretic calibration claim therefore needs either a principled estimator for N, a formal sensitivity analysis, or a statement that the optimal release is conditional on an assumed N.
  5. [Section 2 and Section 3.2.3] The load-bearing assumption that disclosure occurs solely through a certain combination of named entities is contradicted by the authors' own Section 3.2.3 examples: spaCy missed 'iX' in an IBM chunk, and low-concentration clusters paired semantically distant chunks that were still correctly identified after swapping. This is acknowledged as a limitation, but the abstract and conclusion do not carry the qualification. To support corpus-level anonymization guarantees, the method needs a systematic check for chunks whose quasi-identifiers are not captured by the NER step, or the claims must be scoped to 'chunks whose quasi-identifier combinations are successfully extracted and swapped.'
minor comments (4)
  1. [Throughout] There are several typos that should be corrected, including 'metadtata' in Section 2, 'Organizaiton' in Section 3.1.4, 'achives' in Section 3.2.1, and 'speroids' in Section 4.
  2. [Section 3.2.2, Table 3] The McNemar test treats the 30 Monte Carlo draws as independent; because draws share the same underlying corpus and selected pairs, a cluster-robust or draw-level test would be more appropriate, although the observed margins make the conclusion unlikely to change.
  3. [Section 3.2.1] The grid description says '3 x 2^3 = 24' settings but lists only four hyperparameters (N, f, K, epsilon); the breakpoint threshold is discussed separately, so clarify that the grid is over the four listed hyperparameters only.
  4. [Section 3.2.2] The phrase 'the optimal release' is potentially misleading because with r = max chosen for 'most reasonable choices of a and c', the optimization is essentially degenerate; state this explicitly and justify why the maximum achievable r is the relevant operating point.

Circularity Check

1 steps flagged · score 4.0 of 10

Data utility measure is computed under the same fitted spherical mixture that generated the swap weights and constrained swaps to a single cluster, so the utility measure partly certifies its own pairings; external LLM benchmark remains independent.

  1. self definitional [Section 2.5 (Eqs. 2.17-2.19), with Section 2.2.1 and Section 2.3 constraint C]
    "We instead use the likelihood of the mixture model fitted to the pre-swap embedding vector x (which we used in Section 2.2 for the weighting scheme) when constructing the data utility measure DU. In other words, we assume that the mixture model we fitted to pre-swap embedding vector x approximates the true data-generating mechanism of the corpus of text chunks y ∈ Dpre. ... Secondly, the swapping takes place only between chunks that belong to the same cluster in the embedding space."

    The swap weights (Section 2.2.1) and the cluster constraint C (Section 2.3) come from fitting a spherical mixture to pre-swap embeddings x. DU (Eq. 2.19) then scores post-swap data as the log-likelihood ratio of post-swap embeddings under that same fitted mixture, conditional on cluster memberships estimated from that same fitted model. Since every swap is confined to a single fitted cluster, each replacement chunk is drawn from the model's own high-density region, so DU mechanically stays high and partly certifies the model's own pairing choices instead of measuring distortion independently. This biases the risk-utility frontier and the chosen swap rate toward endorsing swaps; the external LLM benchmark in Section 3.2.2 is, however, independent of DU.

full rationale

The principal circular element is the utility measure: Section 2.5 defines DU as the log-likelihood ratio of post-swap to pre-swap embeddings under the very same fitted spherical mixture that Section 2.2.1 used to generate the swap weights and that Section 2.3's cluster constraint uses to restrict all swaps to within-cluster pairs (Eqs. 2.17-2.19). Because every swapped chunk is sourced from the same fitted cluster, the post-swap embeddings remain in the model's high-likelihood region, so DU partly certifies the model's own pairing choices rather than measuring distortion independently. This biases the risk-utility frontier, but it does not by itself carry the paper's central claim: the statistically significant accuracy drop in Table 2 and the McNemar tests are an independent, externally falsifiable LLM benchmark on the swapped chunks. The self-citations (Greve 2025 Eq. 4.11 for the Ewens-Pitman factorial moment; spheroids/Sablica et al. 2025 for the clustering implementation) are not circular in a load-bearing way: the component distributions come from external work (Golzy and Markatou 2020; Kato and McCullagh 2020), the E(S1) identity is standard and already present asymptotically in Hoshino (2001), and the package is independently executable code. Two scope limitations are flagged and weighed: Section 3.2.2 tests only the ~72 swapped chunks (~4% of 1819), leaving the 96% unswapped chunks unexamined, so the abstract's 'prevents' and the conclusion's 'no longer able to guess' overstate corpus-level protection; Section 4 itself concedes 'the swapping can only be performed for a small fraction of the text chunks.' These weaken the scope of the central claim but are not derivation-circularity. Overall score 4.

Assumptions & free parameters 8 free parameters · 6 assumptions · 0 invented entities

The central claim rests on the assumption that named-entity combinations are the sole disclosure channel (Section 2), on the Ewens-Pitman superpopulation model for risk, and on a spherical mixture model for utility that is also used to form swapping pairs. Several hyperparameters are hand-set: N, K, epsilon, f, breakpoint threshold. The mixture parameters and Ewens-Pitman parameters theta and alpha are fitted to data. No new entities are introduced.

free parameters (8)
  • N (superpopulation size for risk table) = grid: 1e10, 1e20, 1e30; final results use 1e20 or 1e30
    Appears in (2.14)-(2.15) for the population-uniques estimate. The ranking of entity-category releases changes with N, and N=1e10 is discarded as unrealistic in Section 3.2.1.
  • K (number of mixture components) = 10 or 30
    Clustering hyperparameter for the swapping weights and utility measure; 30 fails to converge for 6 of 24 settings (Section 3.2.1).
  • epsilon (lower bound on mixture weights) = 1e-3 or 1e-10
    Clustering hyperparameter in spheroids; included in the hyperparameter grid in Section 3.2.1.
  • component density f = PKB or sCauchy
    Choice of component distribution in the spherical mixture used for weights and utility.
  • theta (Ewens-Pitman strength) = 15.4 to 58.9 (Table 1)
    MLE estimate, affects the risk measure through E(S1) in Eq (2.14).
  • alpha (Ewens-Pitman discount) = 0.769 to 0.896 (Table 1)
    MLE estimate, affects the risk measure through E(S1) and p_hat in Eq (2.15).
  • mixture parameters (pi_k, mu_k, rho_k) = estimated by EM; values not reported
    Define the clusters and the utility measure; depend on K, epsilon, and f.
  • breakpoint threshold for semantic chunking = 75 (of 0-100)
    Chunking hyperparameter; authors say the effect on the optimal release is small (Section 3.2.1).
assumptions (6)
  • domain assumption Disclosure of metadata occurs solely through combinations of named entities within text chunks.
    Section 2 states this explicitly and rules out disclosure via style or writing; the whole swapping scheme is built on it.
  • ad hoc to paper The fitted spherical mixture approximates the true data-generating mechanism of the text corpus.
    Section 2.5: 'we assume that the mixture model we fitted to pre-swap embedding vector x approximates the true data-generating mechanism'; used to define DU.
  • domain assumption Ewens-Pitman sampling formula with parameters (theta, alpha) describes the frequency counts of named-entity combinations in the population.
    Section 2.4.2 adopts Hoshino (2001) superpopulation model; the risk measure inherits this assumption.
  • ad hoc to paper Linear risk-utility tradeoff DR = c + aDU.
    Eq (2.8); used to select the optimal release from the efficient frontier.
  • domain assumption Embedding similarity captures semantic interchangeability of named entities.
    Section 2.2.1: clusters of embeddings are treated as candidates for swapping because entities likely serve comparable roles.
  • standard math Singularity analysis asymptotic for the coefficient of (1-t)^{-beta} is valid at the needed large N.
    Used to derive E(S1) ~ N^alpha Gamma(theta+1)/Gamma(theta+alpha) in Eq (2.14).

how reviews work

0 comments
Cite this review

Pith. "Pith review of Named Entity Swapping for Metadata Anonymization in a Text Corpus." pith.science (2026). https://pith.science/paper/3UGM53II

@misc{pith2026250521128,
  author       = {Pith},
  title        = {Pith review of: Named Entity Swapping for Metadata Anonymization in a Text Corpus},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/3UGM53II}},
  note         = {Machine review of arXiv:2505.21128}
}
read the original abstract

This work introduces an anonymization scheme for a corpus of texts to safeguard metadata from disclosure. It specifically aims to prevent large language models from identifying metadata associated with texts, thereby avoiding their influence on query responses. The core mechanism is called named entity swapping, a technique inspired by data swapping in statistical disclosure control. Our method randomly selects pairs of semantically similar substrings from different texts based on the similarity of their embedding vectors and interchanges some named entities between them. This prevents certain combinations of named entities from being uniquely associated with the metadata of individual texts. Our approach offers two key advantages. First, it enables users to determine the optimal level of anonymization that balances data utility and data risk through a calibration of several key decision variables. Second, it leverages text embeddings both to compute swapping weights and to assess data utility, enabling a high degree of flexibility and customization in the overall workflow. The effectiveness of the proposed method is demonstrated with an application that prevents the disclosure of company names in a cross-sectional dataset of earnings call transcripts.

Figures

Figures reproduced from arXiv: 2505.21128 by the authors.

Figure 1
Figure 1. AMD Q4 2014 Earnings Call Opening with Direct Identifiers Suppressed [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. AMD Q4 2014 Earnings Call Opening hyperparameter, which is the aforementioned breakpoint threshold. For our corpus of texts, it produced 1819 chunks in total. Hence, n is set to 1819 in the subsequent analysis. 3.1.3 Suppression of Direct Identifiers Direct identifiers are information that uniquely maps to the sensitive metadata we aim to protect from disclosure. All direct identifiers that reveal the company name s… view at source ↗
Figure 3
Figure 3. Risk-Utility Plots for PKB Mixtures with [PITH_FULL_IMAGE:figures/full_fig_p016_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Risk-Utility Plots for the Rest of the PKB Mixtures [PITH_FULL_IMAGE:figures/full_fig_p024_4.png]
Figure 5
Figure 5. Figure 5: Risk-Utility Plots for the Rest of the Spherical Cauchy Mixtures [PITH_FULL_IMAGE:figures/full_fig_p025_5.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

32 extracted references · 24 canonical work pages

  1. [1]

    Arbel, S

    J. Arbel, S. Favaro, B. Nipoti, and Y. W. Teh. Bayesian nonparametric inference for discovery probabilities: Credible intervals and large sample asymptiotics . Statistica Sinica, pages 839--858, 2017

    work page 2017

  2. [2]

    L. P. Argyle, E. C. Busby, N. Fulda, J. R. Gubler, C. Rytting, and D. Wingate. Out of One, Many: Using Language Models to Simulate Human Samples . Political Analysis, 31 0 (3): 0 337--351, 2023

    work page 2023

  3. [3]

    Chen and X

    J. Chen and X. Tan. Inference for multivariate normal mixtures. Journal of Multivariate Analysis, 100 0 (7): 0 1367--1383, 2009

    work page 2009

  4. [4]

    J. Chen, X. Tan, and R. Zhang. Inference for normal mixtures in mean and variance. Statistica Sinica, 18: 0 443--465, 2008

    work page 2008

  5. [5]

    Dalenius and S

    T. Dalenius and S. P. Reiss. Data-swapping: A technique for disclosure control . Journal of Statistical Planning and Inference, 6 0 (1): 0 73--85, 1982

    work page 1982

  6. [6]

    A. P. Dempster, N. M. Laird, and D. B. Rubin. Maximum likelihood from incomplete data via the em algorithm. Journal of the Royal Statistical Society: Series B (Methodological), 39 0 (1): 0 1--38, 1977

    1977

  7. [7]

    Efron and R

    B. Efron and R. Thisted. Estimating the Number of Unseen Species: How Many Words Did Shakespeare Know? Biometrika, 63 0 (3): 0 435--447, 1976

    work page 1976

  8. [8]

    Favaro, B

    S. Favaro, B. Nipoti, and Y. W. Teh. Rediscovery of Good--Turing estimators via Bayesian nonparametrics . Biometrics, 72 0 (1): 0 136--145, 2016

    work page 2016

Show all 32 references
  1. [9]

    S. E. Fienberg and J. McIntyre. Data Swapping: Variations on a Theme by Dalenius and Reiss . In Privacy in Statistical Databases, pages 14--29, 2004

    2004

  2. [10]

    Flajolet and A

    P. Flajolet and A. Odlyzko. Singularity analysis of generating functions. SIAM Journal on Discrete Mathematics, 3 0 (2): 0 216--240, 1990

    1990

  3. [11]

    Gentzkow, B

    M. Gentzkow, B. Kelly, and M. Taddy. Text as Data . Journal of Economic Literature, 57 0 (3): 0 535–74, 2019

    2019

  4. [12]

    Glasserman and C

    P. Glasserman and C. Lin. Assessing Look-Ahead Bias in Stock Return Predictions Generated By GPT Sentiment Analysis . arXiv preprint arXiv:2309.17322, 2023

    2023 arXiv

  5. [13]

    Golzy and M

    M. Golzy and M. Markatou. Poisson Kernel-Based Clustering on the Sphere: Convergence Properties, Identifiability, and a Method of Sampling . Journal of Computational and Graphical Statistics, 29 0 (4): 0 758--770, 2020

    2020

  6. [14]

    Gomatam, A

    S. Gomatam, A. F. Karr, and A. P. Sanil. Data Swapping as a Decision Problem . Journal of Official Statistics, 21 0 (4): 0 635, 2005

    2005

  7. [15]

    J. Greve. A New Representation of Ewens-Pitman's Partition Structure and Its Characterization via Riordan Array Sums . arXiv preprint arXiv:2503.05034, 2025

    2025 arXiv

  8. [16]

    Halawi, F

    D. Halawi, F. Zhang, C. Yueh-Han, and J. Steinhardt. Approaching Human-Level Forecasting with Language Models . arXiv preprint arXiv:2402.18563, 2024

    2024 arXiv

  9. [17]

    Honnibal, I

    M. Honnibal, I. Montani, S. V. Landeghem, and A. Boyd. spaCy: Industrial-strength Natural Language Processing in Python . 2020. doi:10.5281/zenodo.1212303

    2020 doi

  10. [18]

    J. J. Horton. Large Language Models as Simulated Economic Agents: What Can We Learn from Homo Silicus? Technical report, 2023

    2023

  11. [19]

    N. Hoshino. Applying Pitman's Sampling Formula to Microdata Disclosure Risk Assessment . Journal of Official Statistics, 17 0 (4): 0 499--520, 2001

    2001

  12. [20]

    Jurafsky and J

    D. Jurafsky and J. H. Martin. Speech and Language Processing: An Introduction to Natural Language Processing, Computational Linguistics, and Speech Recognition. Prentice Hall PTR, 1 edition, 2000. ISBN 0130950696

    2000

  13. [21]

    Kato and P

    S. Kato and P. McCullagh. Some properties of a Cauchy family on the sphere derived from the M \"o bius transformations . Bernoulli, 26: 0 3224--3248, 2020

    2020

  14. [22]

    Khatri and K

    C. Khatri and K. V. Mardia. The von Mises--Fisher Matrix Distribution in Orientation Statistics . Journal of the Royal Statistical Society Series B (Methodological), 39 0 (1): 0 95--106, 1977

    1977

  15. [23]

    Semantic Chunking Example — LlamaIndex Documentation , 2024

    LlamaIndex . Semantic Chunking Example — LlamaIndex Documentation , 2024. URL https://docs.llamaindex.ai/en/stable/examples/node_parsers/semantic_chunking/. Accessed: 2025-04-07

    2024

  16. [24]

    Lopez-Lira and Y

    A. Lopez-Lira and Y. Tang. Can ChatGPT Forecast Stock Price Movements? Return Predictability and Large Language Models . arXiv preprint arXiv:2304.07619, 2023

    2023

  17. [25]

    text-embedding-3-small

    OpenAI. text-embedding-3-small. https://platform.openai.com/docs/guides/embeddings, 2023. Accessed: 2025-04-22

    2023

  18. [26]

    L. Sablica. spheroids: A Python Package for Spherical Clustering Models , 2025. URL https://github.com/lsablica/spheroids

    2025

  19. [27]

    Sablica, K

    L. Sablica, K. Hornik, and B. Gr \"u n. circlus: An R Package for Circular and Spherical Clustering using Poisson Kernel-Based and Spherical Cauchy Distributions . To appear in Austrian Journal os Statistics, 2025

    2025

  20. [28]

    S. M. Samuels. A Bayesian, species-sampling-inspired approach to the uniques problem in microdata disclosure risk assessment . Journal of Official Statistics, 14 0 (4): 0 373, 1998

    1998

  21. [29]

    S. K. Sarkar and K. Vafa. Lookahead bias in pretrained language models. Sarkar, Suproteem and Vafa, Keyon, Lookahead Bias in Pretrained Language Models (June 28, 2024). Available at SSRN: https://ssrn.com/abstract=4754678 or http://dx.doi.org/10.2139/ssrn.4754678, 2024

    2024 doi

  22. [30]

    H. Su, W. Shi, J. Kasai, Y. Wang, Y. Hu, M. Ostendorf, W.-t. Yih, N. A. Smith, L. Zettlemoyer, and T. Yu. One Embedder, Any Task: Instruction-Finetuned Text Embeddings . arXiv preprint arXiv:2212.09741, 2022

    2022 arXiv

  23. [31]

    Takemura

    A. Takemura. Local Recoding and Record Swapping by Maximum Weight Matching for Disclosure Control of Microdata Sets . Journal of Official Statistics, 18 0 (2): 0 275--289, 2002

    2002

  24. [32]

    Vaswani, N

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, . Kaiser, and I. Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017

    2017

Pith tools

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